阅读说明：本技术 资源推荐模型的训练方法、资源推荐方法、装置及服务器 (Resource recommendation model training method, resource recommendation device and server ) 是由 廖一桥 于 2021-08-27 设计创作，主要内容包括：本公开关于一种资源推荐模型的训练方法、资源推荐方法、装置及服务器,属于人工智能技术领域。方法包括：从多个资源类别中,确定新增行为特征对应的第一资源类别,新增行为特征是基于新增行为数据确定的；确定第一资源类别对应的第一类别特征,将新增行为特征与第一类别特征进行融合,得到第一资源类别更新后的第一类别特征；基于多个资源类别对应的类别特征,确定资源推荐模型的第一样本特征,多个资源类别对应的类别特征包括第一资源类别更新后的第一类别特征,基于第一样本特征训练资源推荐模型,资源推荐模型用于向任一用户账号推荐资源。上述方案在提高了训练资源推荐模型的准确率的同时,还降低了存储压力和处理压力。(The disclosure relates to a resource recommendation model training method, a resource recommendation device and a server, and belongs to the technical field of artificial intelligence. The method comprises the following steps: determining a first resource category corresponding to the newly added behavior characteristics from the plurality of resource categories, wherein the newly added behavior characteristics are determined based on the newly added behavior data; determining a first class feature corresponding to the first resource class, and fusing the newly added behavior feature with the first class feature to obtain the updated first class feature of the first resource class; the resource recommendation method includes the steps of determining first sample characteristics of a resource recommendation model based on category characteristics corresponding to multiple resource categories, wherein the category characteristics corresponding to the multiple resource categories comprise the first category characteristics after the first resource categories are updated, training the resource recommendation model based on the first sample characteristics, and the resource recommendation model is used for recommending resources to any user account. According to the scheme, the accuracy of the resource recommendation model is improved, and meanwhile, the storage pressure and the processing pressure are reduced.)

1. A method for training a resource recommendation model is characterized by comprising the following steps:

determining a first resource category corresponding to a newly added behavior feature from a plurality of resource categories, wherein the newly added behavior feature is determined based on newly added behavior data, the newly added behavior data is generated based on the operation of a user account on a resource, and the first resource category is the category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data;

determining a first category feature corresponding to the first resource category, and fusing the newly added behavior feature and the first category feature to obtain the updated first category feature of the first resource category;

determining first sample characteristics of a resource recommendation model based on the category characteristics corresponding to the resource categories, wherein the category characteristics corresponding to the resource categories comprise the updated first category characteristics of the first resource category, training the resource recommendation model based on the first sample characteristics, and the resource recommendation model is used for recommending resources to any user account.

2. The method of claim 1, wherein the class characteristics corresponding to each resource class are stored in a bucket corresponding to the resource class, and the merging the new behavior characteristics with the first class characteristics to obtain the updated first class characteristics of the first resource class comprises:

determining a first bucket corresponding to the first resource category from buckets corresponding to the plurality of resource categories;

extracting the first class feature from the first bucket;

and fusing the newly added behavior features and the first class features to obtain the updated first class features of the first resource class, and storing the updated first class features in the first bucket.

3. The method of claim 2, wherein the added behavior feature comprises a plurality of dimensions of added sub-features;

each resource category corresponds to a different bucket, each bucket corresponds to a different dimension, and each bucket is used for storing a category sub-feature belonging to the dimension corresponding to the bucket under the resource category; and the category characteristics corresponding to each resource category comprise category sub-characteristics of the multiple dimensions;

the merging the newly added behavior features with the first category features to obtain the updated first category features of the first resource category, and storing the updated first category features in the first bucket includes:

for each dimension of the plurality of dimensions, determining a second bucket corresponding to the dimension from the first bucket, and extracting category sub-features from the second bucket;

and fusing the newly added sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in the second bucket.

4. The method according to any one of claims 1-3, further comprising:

acquiring configuration information, wherein the configuration information comprises resource categories of n hierarchies, the resource categories are resource categories of the nth hierarchy in the n hierarchies, and n is a positive integer greater than 1;

the determining a first sample characteristic of a resource recommendation model based on the category characteristics corresponding to the plurality of resource categories includes:

in the ith level, fusing category characteristics corresponding to a plurality of resource categories of the current level belonging to the same superior resource category, and determining the fused category characteristics as the category characteristics corresponding to the superior resource category until the category characteristics corresponding to at least one resource category in the kth level are obtained, wherein i and k are any positive integer not greater than n, i is greater than k, and the resource category in the (i-1) level is the superior resource category of the resource category in the ith level;

and determining a category characteristic corresponding to at least one resource category in the kth level as the first sample characteristic.

5. A method for resource recommendation, the method comprising:

determining resource characteristics of resources to be recommended and user characteristics of a target user account;

determining model input features based on the resource features, the user features and category features corresponding to a plurality of resource categories, wherein each category feature is obtained by respectively fusing historical behavior features corresponding to each resource category, and the historical behavior features are determined according to historical behavior data;

inputting the model input characteristics into a resource recommendation model to obtain a prediction recommendation parameter corresponding to the resource to be recommended;

and recommending the resources to be recommended to the target user account based on the predicted recommendation parameters.

6. An apparatus for training a resource recommendation model, comprising:

the resource management system comprises a first determining unit, a second determining unit and a resource management unit, wherein the first determining unit is configured to determine a first resource category corresponding to a newly added behavior characteristic from a plurality of resource categories, the newly added behavior characteristic is determined based on newly added behavior data, the newly added behavior data is generated based on an operation of a user account on a resource, and the first resource category is a category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data;

the fusion unit is configured to execute the first class feature corresponding to the first resource class, fuse the newly added behavior feature with the first class feature, and obtain the updated first class feature of the first resource class;

the training unit is configured to execute determining a first sample feature of a resource recommendation model based on category features corresponding to the multiple resource categories, wherein the category features corresponding to the multiple resource categories include the updated first category feature of the first resource category, and train the resource recommendation model based on the first sample feature, and the resource recommendation model is used for recommending resources to any user account.

7. An apparatus for resource recommendation, the apparatus comprising:

the third determination unit is configured to perform determination of the resource characteristics of the resource to be recommended and the user characteristics of the target user account;

a fourth determining unit, configured to perform determining model input features based on the resource features, the user features and category features currently corresponding to multiple resource categories, where each category feature is obtained by fusing historical behavior features corresponding to each resource category respectively, and the historical behavior features are determined according to historical behavior data;

a fifth determining unit, configured to input the model input features into a resource recommendation model, so as to obtain a prediction recommendation parameter corresponding to the resource to be recommended;

and the recommending unit is configured to recommend the resource to be recommended to the target user account based on the predicted recommending parameter.

8. A server, comprising:

a processor;

a memory for storing the processor-executable instructions;

wherein the processor is configured to execute the instructions to implement the method of training of the resource recommendation model according to any one of claims 1 to 4 or the method of resource recommendation according to claim 5.

9. A computer-readable storage medium, wherein instructions in the computer-readable storage medium, when executed by a processor of a server, enable the server to perform the method of training a resource recommendation model according to any one of claims 1 to 4 or the method of resource recommendation according to claim 5.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the method of training a resource recommendation model according to any one of claims 1 to 4 or the method of resource recommendation according to claim 5.

Technical Field

The disclosure relates to the technical field of artificial intelligence, in particular to a resource recommendation model training method, a resource recommendation device and a server.

Background

With the development of the internet, a plurality of resource display applications are generated, and each resource display application can recommend resources which are interested in the resource display application for a user so as to improve the viscosity of the user. In order to improve the recommendation accuracy, resource recommendation is usually performed based on a resource recommendation model, which requires to train an accurate resource recommendation model first.

The behavior characteristics have important guiding significance for training of the resource recommendation model, are generated according to the behavior data of the user, can represent the behavior of the user on the resources, and can reflect the interest of the user on the resources. The more behavior features are used in training the resource recommendation model, the higher the accuracy of the resource recommendation model is, but a large number of behavior features need to be stored, and the large number of behavior features are read in training, which causes large storage pressure and processing pressure. Therefore, how to reduce the storage pressure and the processing pressure while training the resource recommendation model according to more behavior characteristics becomes an urgent problem to be solved.

Disclosure of Invention

The disclosure provides a resource recommendation model training method, a resource recommendation device and a server, which are used for improving the accuracy of resource recommendation model training and reducing the storage pressure and the processing pressure. The technical scheme of the disclosure is as follows:

according to a first aspect of the embodiments of the present disclosure, a method for training a resource recommendation model is provided, including:

determining a first resource category corresponding to a newly added behavior feature from a plurality of resource categories, wherein the newly added behavior feature is determined based on newly added behavior data, the newly added behavior data is generated based on the operation of a user account on a resource, and the first resource category is the category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data;

determining a first category feature corresponding to the first resource category, and fusing the newly added behavior feature and the first category feature to obtain the updated first category feature of the first resource category;

determining first sample characteristics of a resource recommendation model based on the category characteristics corresponding to the resource categories, wherein the category characteristics corresponding to the resource categories comprise the updated first category characteristics of the first resource category, training the resource recommendation model based on the first sample characteristics, and the resource recommendation model is used for recommending resources to any user account.

In some embodiments, the storing the category characteristics corresponding to each resource category in the bucket corresponding to the resource category, and the merging the newly added behavior characteristics with the first category characteristics to obtain the updated first category characteristics of the first resource category includes:

determining a first bucket corresponding to the first resource category from buckets corresponding to the plurality of resource categories;

extracting the first class feature from the first bucket;

and fusing the newly added behavior features and the first class features to obtain the updated first class features of the first resource class, and storing the updated first class features in the first bucket.

In some embodiments, the added behavior feature comprises a plurality of dimensions of added sub-features;

each resource category corresponds to a different bucket, each bucket corresponds to a different dimension, and each bucket is used for storing a category sub-feature belonging to the dimension corresponding to the bucket under the resource category; and the category characteristics corresponding to each resource category comprise category sub-characteristics of the multiple dimensions;

the merging the newly added behavior features with the first category features to obtain the updated first category features of the first resource category, and storing the updated first category features in the first bucket includes:

for each dimension of the plurality of dimensions, determining a second bucket corresponding to the dimension from the first bucket, and extracting category sub-features from the second bucket;

and fusing the newly added sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in the second bucket.

In some embodiments, the determining, based on the category characteristics corresponding to the plurality of resource categories, a first sample characteristic of a resource recommendation model includes:

for each resource category, respectively extracting category sub-features corresponding to the resource category from a plurality of buckets corresponding to the resource category;

and splicing the extracted category sub-features corresponding to the plurality of resource categories to obtain the first sample feature.

In some embodiments, the method further comprises:

obtaining sample behavior data, wherein the sample behavior data comprises resource features, user features and behavior labels of recommended sample resources, the sample resources are resources recommended to a sample user account after category features corresponding to the multiple resource categories are obtained, the resource features are used for describing the sample resources, the user features are user features corresponding to the sample user account, and the behavior labels are used for representing types of behaviors of the sample user account on the sample resources;

determining the resource characteristics and the user characteristics of the sample resources as second sample characteristics;

the training the resource recommendation model based on the first sample features includes:

training the resource recommendation model based on the first sample features, the second sample features, and the behavior labels.

In some embodiments, the training the resource recommendation model based on the first sample feature, the second sample feature, and the behavior label includes:

splicing the first sample characteristic and the second sample characteristic to obtain a third sample characteristic, and training the resource recommendation model based on the third sample characteristic and the behavior label; alternatively, the first and second electrodes may be,

and fusing the first sample characteristic and the second sample characteristic to obtain a third sample characteristic, and training the resource recommendation model based on the third sample characteristic and the behavior label.

In some embodiments, the merging the new-addition behavior feature with the first class feature to obtain the updated first class feature of the first resource class includes:

determining the newly added behavior characteristics distributed to each first resource type according to the product of the distribution proportion corresponding to each first resource type and the newly added behavior characteristics;

and for each first resource category, fusing the newly added behavior characteristics obtained by the allocation of the first resource category with the first category characteristics corresponding to the first resource category to obtain the updated first category characteristics of the first resource category.

In some embodiments, the method further comprises:

acquiring configuration information, wherein the configuration information comprises resource categories of n hierarchies, the resource categories are resource categories of the nth hierarchy in the n hierarchies, and n is a positive integer greater than 1;

the determining a first sample characteristic of a resource recommendation model based on the category characteristics corresponding to the plurality of resource categories includes:

in the ith level, fusing category characteristics corresponding to a plurality of resource categories of the current level belonging to the same superior resource category, and determining the fused category characteristics as the category characteristics corresponding to the superior resource category until the category characteristics corresponding to at least one resource category in the kth level are obtained, wherein i and k are any positive integer not greater than n, i is greater than k, and the resource category in the (i-1) level is the superior resource category of the resource category in the ith level;

and determining a category characteristic corresponding to at least one resource category in the kth level as the first sample characteristic.

In some embodiments, the determining, from the plurality of resource categories, a first resource category corresponding to the new added behavior feature includes:

acquiring the newly added behavior data, wherein the newly added behavior data at least comprises at least one item of user data and at least one item of resource data, and the at least one item of resource data comprises the first resource type;

and extracting the characteristics of the data except the first resource type in the newly added behavior data to obtain the newly added behavior characteristics, and determining the first resource type as the first resource type corresponding to the newly added behavior characteristics.

In some embodiments, the merging the new behavior feature with the first category feature to obtain the updated first category feature of the first resource category includes:

and according to the weight of the newly added behavior feature and the first class feature, performing weighted fusion on the newly added behavior feature and the first class feature to obtain the first class feature after the first resource class is updated.

According to a second aspect of the embodiments of the present disclosure, there is provided a resource recommendation method, the method including:

determining resource characteristics of resources to be recommended and user characteristics of a target user account;

determining model input features based on the resource features, the user features and category features corresponding to a plurality of resource categories, wherein each category feature is obtained by respectively fusing historical behavior features corresponding to each resource category, and the historical behavior features are determined according to historical behavior data;

inputting the model input characteristics into a resource recommendation model to obtain a prediction recommendation parameter corresponding to the resource to be recommended;

and recommending the resources to be recommended to the target user account based on the predicted recommendation parameters.

In some embodiments, the method further comprises:

acquiring configuration information, wherein the configuration information comprises resource categories of n hierarchies, the resource categories are resource categories of the nth hierarchy in the n hierarchies, and n is a positive integer greater than 1;

determining a model input feature based on the resource feature, the user feature, and category features corresponding to a plurality of resource categories, comprising:

and determining the model input characteristics based on the resource characteristics, the user characteristics and the category characteristics corresponding to at least one resource category in the kth level, wherein k is a positive integer not larger than n.

According to a third aspect of the embodiments of the present disclosure, there is provided a training apparatus for a resource recommendation model, including:

the resource management system comprises a first determining unit, a second determining unit and a resource management unit, wherein the first determining unit is configured to determine a first resource category corresponding to a newly added behavior characteristic from a plurality of resource categories, the newly added behavior characteristic is determined based on newly added behavior data, the newly added behavior data is generated based on an operation of a user account on a resource, and the first resource category is a category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data;

the fusion unit is configured to execute the first class feature corresponding to the first resource class, fuse the newly added behavior feature with the first class feature, and obtain the updated first class feature of the first resource class;

the training unit is configured to execute determining a first sample feature of a resource recommendation model based on category features corresponding to the multiple resource categories, wherein the category features corresponding to the multiple resource categories include the updated first category feature of the first resource category, and train the resource recommendation model based on the first sample feature, and the resource recommendation model is used for recommending resources to any user account.

In some embodiments, the class feature corresponding to each resource class is stored in a bucket corresponding to the resource class, and the merging unit includes:

a determining subunit configured to perform determining a first bucket corresponding to the first resource category from among buckets corresponding to the plurality of resource categories;

an extraction subunit configured to perform extraction of the first class feature from the first bucket;

and the fusion subunit is configured to perform fusion of the newly added behavior feature and the first class feature to obtain an updated first class feature of the first resource class, and store the updated first class feature in the first bucket.

In some embodiments, the added behavior feature comprises a plurality of dimensions of added sub-features;

each resource category corresponds to a different bucket, each bucket corresponds to a different dimension, and each bucket is used for storing a category sub-feature belonging to the dimension corresponding to the bucket under the resource category; and the category characteristics corresponding to each resource category comprise category sub-characteristics of the multiple dimensions;

the fusion subunit is configured to determine, for each of the plurality of dimensions, a second bucket corresponding to the dimension from the first bucket, and extract a category sub-feature from the second bucket; and fusing the newly added sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in the second bucket.

In some embodiments, the training unit is configured to perform, for each resource category, extracting category sub-features corresponding to the resource category from a plurality of buckets corresponding to the resource category, respectively; and splicing the extracted category sub-features corresponding to the plurality of resource categories to obtain the first sample feature.

In some embodiments, the apparatus further comprises:

a first obtaining unit, configured to perform obtaining of sample behavior data, where the sample behavior data includes resource features, user features, and behavior tags of recommended sample resources, where the sample resources are resources recommended to a sample user account after category features corresponding to the multiple resource categories are obtained, the resource features are used for describing the sample resources, the user features are user features corresponding to the sample user account, and the behavior tags are used for representing types of behaviors performed by the sample user account on the sample resources;

a second determination unit configured to perform determination of the resource feature and the user feature of the sample resource as a second sample feature;

the training unit is configured to perform training of the resource recommendation model based on the first sample feature, the second sample feature and the behavior label.

In some embodiments, the training unit is configured to perform stitching of the first sample feature and the second sample feature to obtain a third sample feature, and train the resource recommendation model based on the third sample feature and the behavior label; alternatively, the first and second electrodes may be,

the training unit is configured to perform fusion of the first sample feature and the second sample feature to obtain a third sample feature, and train the resource recommendation model based on the third sample feature and the behavior label.

In some embodiments, the resource belongs to a plurality of first resource categories, and the merging unit is configured to perform a multiplication of an allocation proportion corresponding to each first resource category and the new-added behavior feature, and determine a new-added behavior feature allocated to each first resource category; and for each first resource category, fusing the newly added behavior characteristics obtained by the allocation of the first resource category with the first category characteristics corresponding to the first resource category to obtain the updated first category characteristics of the first resource category.

In some embodiments, the apparatus further comprises:

a second obtaining unit configured to perform obtaining configuration information, where the configuration information includes resource categories of n hierarchies, where the plurality of resource categories are resource categories of an nth level in the n hierarchies, and n is a positive integer greater than 1;

the training unit is configured to perform fusion of category features corresponding to a plurality of resource categories of a current level belonging to the same superior resource category in an ith level, and determine the fused category features as category features corresponding to the superior resource category until a category feature corresponding to at least one resource category in a kth level is obtained, where i and k are any positive integer not greater than n, i is greater than k, and the resource category in the (i-1) level is the superior resource category of the resource category in the ith level; and determining a category characteristic corresponding to at least one resource category in the kth level as the first sample characteristic.

In some embodiments, the first determining unit is configured to perform obtaining the new adding behavior data, where the new adding behavior data includes at least one item of user data and at least one item of resource data, and the at least one item of resource data includes the first resource category; and extracting the characteristics of the data except the first resource type in the newly added behavior data to obtain the newly added behavior characteristics, and determining the first resource type as the first resource type corresponding to the newly added behavior characteristics.

In some embodiments, the merging unit is configured to perform weighted merging on the new added behavior feature and the first category feature according to the weights of the new added behavior feature and the first category feature, so as to obtain the updated first category feature of the first resource category.

According to a fourth aspect of the embodiments of the present disclosure, there is provided a resource recommendation apparatus, the apparatus including:

the third determination unit is configured to perform determination of the resource characteristics of the resource to be recommended and the user characteristics of the target user account;

a fourth determining unit, configured to perform determining model input features based on the resource features, the user features and category features currently corresponding to multiple resource categories, where each category feature is obtained by fusing historical behavior features corresponding to each resource category respectively, and the historical behavior features are determined according to historical behavior data;

a fifth determining unit, configured to input the model input features into a resource recommendation model, so as to obtain a prediction recommendation parameter corresponding to the resource to be recommended;

and the recommending unit is configured to recommend the resource to be recommended to the target user account based on the predicted recommending parameter.

In some embodiments, the apparatus further comprises:

a third obtaining unit configured to perform obtaining configuration information, where the configuration information includes resource categories of n hierarchies, where the plurality of resource categories are resource categories of an nth level in the n hierarchies, and n is a positive integer greater than 1;

the fourth determining unit is configured to perform determining the model input feature based on the resource feature, the user feature and a category feature corresponding to at least one resource category in a kth level, where k is a positive integer not greater than n.

According to a fifth aspect of embodiments of the present disclosure, there is provided a server including:

a processor;

a memory for storing the processor-executable instructions;

wherein the processor is configured to execute the instructions to implement the method for training a resource recommendation model according to the first aspect or the method for resource recommendation according to the second aspect.

According to a sixth aspect of embodiments of the present disclosure, there is provided a computer-readable storage medium, wherein instructions, when executed by a processor of a server, enable the server to perform the method of training a resource recommendation model according to the first aspect or the method of resource recommendation according to the second aspect.

According to a seventh aspect of embodiments of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the method for training a resource recommendation model according to the first aspect or the method for resource recommendation according to the second aspect.

In the resource recommendation model training method, the resource recommendation device and the server provided by the embodiments of the present disclosure, each newly added behavior feature is not stored, but the newly added behavior features are fused with the corresponding stored category features according to the resource category, so that the updated category features fuse information included in the newly added behavior features, the resource recommendation model can learn the newly added behavior features during training according to the updated category features, and the newly added behavior features are not required to be stored again because the updated category features are fused with the newly added behavior features, and further, the above-mentioned scheme improves the accuracy of the resource recommendation model training and reduces the storage pressure and the processing pressure.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the disclosure and are not to be construed as limiting the disclosure.

FIG. 1 is a schematic diagram of an implementation environment shown in accordance with an exemplary embodiment;

FIG. 2 is a flow diagram illustrating a method of training a resource recommendation model in accordance with an exemplary embodiment;

FIG. 3 is a flow diagram illustrating a method for resource recommendation in accordance with an exemplary embodiment;

FIG. 4 is a flow diagram illustrating a method of training a resource recommendation model in accordance with an exemplary embodiment;

FIG. 5 is a schematic diagram illustrating a method of training a resource recommendation model in accordance with an exemplary embodiment;

FIG. 6 is a flow diagram illustrating a method of training a resource recommendation model in accordance with an exemplary embodiment;

FIG. 7 is a flow diagram illustrating a method for resource recommendation in accordance with an exemplary embodiment;

FIG. 8 is a block diagram illustrating an architecture of a training apparatus for a resource recommendation model in accordance with an exemplary embodiment;

FIG. 9 is a block diagram illustrating an architecture of a training apparatus for a resource recommendation model in accordance with an exemplary embodiment;

FIG. 10 is a block diagram illustrating the structure of a resource recommendation device in accordance with an exemplary embodiment;

FIG. 11 is a block diagram illustrating the structure of a resource recommendation device in accordance with an exemplary embodiment;

FIG. 12 is a block diagram illustrating the structure of a server in accordance with an exemplary embodiment.

Detailed Description

In order to make the technical solutions of the present disclosure better understood by those of ordinary skill in the art, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first," "second," and the like in the description and claims of the present disclosure and in the above-described drawings are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the disclosure described herein are capable of operation in sequences other than those illustrated or otherwise described herein. The implementations described in the exemplary embodiments below are not intended to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.

The user information to which the present disclosure relates may be information authorized by the user or sufficiently authorized by each party.

FIG. 1 is a schematic diagram illustrating one implementation environment in accordance with an example embodiment. Referring to fig. 1, the implementation environment includes a terminal 101 and a server 102. The terminal 101 and the server 102 are connected via a wireless or wired network.

Illustratively, the terminal 101 has installed thereon an application served by the server 102, and the terminal 101 can implement functions such as resource presentation or data transmission through the application. Illustratively, the application is an application in an operating system of the terminal 101 or an application provided by a third party. The application has a resource presentation function, for example, the application has a function of presenting a video, presenting an image, and presenting an audio. Of course, the application can also have other functions, such as a game function, a shopping function, a chat function, and the like. Illustratively, the application is a short video application, a gaming application, a shopping application, an audio application, or other applications, which are not limited by the embodiments of the present disclosure.

In the embodiment of the present disclosure, a user performs one or more behaviors on a resource on the terminal 101, such as a praise behavior, a comment behavior, and the like, the terminal 101 acquires behavior data, sends the behavior data to the server 102, and the server 102 processes the behavior data to obtain a behavior characteristic. By adopting the above manner, the server 102 may obtain one or more behavior characteristics, so that the resource recommendation model can be trained according to the obtained behavior characteristics, and resources are recommended to the terminal 101 through the trained resource recommendation model.

The resource recommendation model training method provided by the embodiment of the disclosure is applied to various scenes.

For example, taking a resource as a video as an example, when a user browses the video through an application installed in a terminal, a behavior of the user on the video is recorded as new behavior data. By adopting the method provided by the embodiment of the disclosure, the newly added behavior feature is generated according to the newly added behavior data, and the acquired newly added behavior feature is fused with the stored category feature according to the resource category, so that the resource recommendation model can learn more newly added behavior features, and the storage pressure and the processing pressure of a large amount of newly added behavior features are reduced. And subsequently, when the resource recommendation is performed on the terminal, the resource recommended for the terminal can be determined according to the trained resource recommendation model.

FIG. 2 is a flowchart illustrating a method of training a resource recommendation model, as shown in FIG. 2, including the following steps, according to an example embodiment.

201. Determining a first resource category corresponding to the newly added behavior feature from the multiple resource categories, wherein the newly added behavior feature is determined based on newly added behavior data, the newly added behavior data is generated based on the operation of a user account on the resource, and the first resource category is the category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing the historical new-adding behavior characteristics corresponding to each resource category, and the historical new-adding behavior characteristics are determined according to historical behavior data.

202. And determining a first class feature corresponding to the first resource class, and fusing the newly added behavior feature with the first class feature to obtain the updated first class feature of the first resource class.

203. The resource recommendation method includes the steps of determining first sample characteristics of a resource recommendation model based on category characteristics corresponding to multiple resource categories, wherein the category characteristics corresponding to the multiple resource categories comprise the first category characteristics after the first resource categories are updated, training the resource recommendation model based on the first sample characteristics, and the resource recommendation model is used for recommending resources to any user account.

In some embodiments, the storing the category characteristics corresponding to each resource category in a bucket corresponding to the resource category, and fusing the new behavior characteristics with the first category characteristics to obtain the updated first category characteristics of the first resource category includes:

determining a first bucket corresponding to a first resource category from buckets corresponding to a plurality of resource categories;

extracting a first class feature from the first bucket;

and fusing the newly added behavior features and the first class features to obtain the updated first class features of the first resource class, and storing the updated first class features in the first bucket.

In some embodiments, the added behavior feature comprises a plurality of dimensions of added sub-features;

each resource category corresponds to different buckets, each bucket corresponds to different dimensionality, and each bucket is used for storing a category sub-feature belonging to the corresponding dimensionality of the bucket under the resource category; the category characteristics corresponding to each resource category comprise category sub-characteristics of a plurality of dimensions;

the newly added behavior features and the first class features are fused to obtain the first class features after the first resource class is updated, and the updated first class features are stored in a first bucket, wherein the method comprises the following steps:

for each dimension in the plurality of dimensions, determining a second bucket corresponding to the dimension from the first bucket, and extracting category sub-features from the second bucket;

and fusing the new increasing sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in a second bucket.

In some embodiments, determining a first sample feature of the resource recommendation model based on category features corresponding to a plurality of resource categories includes:

for each resource category, extracting category sub-features corresponding to the resource category from a plurality of buckets corresponding to the resource category respectively;

and splicing the extracted category sub-features corresponding to the plurality of resource categories to obtain a first sample feature.

In some embodiments, the method further comprises:

obtaining sample behavior data, wherein the sample behavior data comprises resource characteristics, user characteristics and behavior labels of recommended sample resources, the sample resources are resources recommended to a sample user account after category characteristics corresponding to a plurality of resource categories are obtained, the resource characteristics are used for describing the sample resources, the user characteristics are user characteristics corresponding to the sample user account, and the behavior labels are used for expressing types of behaviors of the sample user account on the sample resources;

determining the resource characteristics and the user characteristics of the sample resources as second sample characteristics;

training a resource recommendation model based on the first sample features, comprising:

and training a resource recommendation model based on the first sample characteristic, the second sample characteristic and the behavior label.

In some embodiments, training the resource recommendation model based on the first sample features, the second sample features, and the behavior labels includes:

splicing the first sample characteristic and the second sample characteristic to obtain a third sample characteristic, and training a resource recommendation model based on the third sample characteristic and the behavior label; alternatively, the first and second electrodes may be,

and fusing the first sample characteristic and the second sample characteristic to obtain a third sample characteristic, and training a resource recommendation model based on the third sample characteristic and the behavior label.

In some embodiments, the method for merging the new adding behavior feature with the first category feature to obtain the updated first category feature of the first resource category includes:

determining the newly added behavior characteristics distributed for each first resource type by respectively multiplying the distribution proportion corresponding to each first resource type by the newly added behavior characteristics;

and for each first resource category, fusing the newly added behavior characteristics obtained by the allocation of the first resource category with the first category characteristics corresponding to the first resource category to obtain the first category characteristics after the updating of the first resource category.

In some embodiments, the method further comprises:

acquiring configuration information, wherein the configuration information comprises resource categories of n hierarchies, a plurality of resource categories are resource categories of the nth hierarchy in the n hierarchies, and n is a positive integer greater than 1;

determining a first sample feature of a resource recommendation model based on category features corresponding to a plurality of resource categories, including:

in the ith level, fusing category characteristics corresponding to a plurality of resource categories of the current level belonging to the same superior resource category, and determining the fused category characteristics as the category characteristics corresponding to the superior resource category until the category characteristics corresponding to at least one resource category in the kth level are obtained, wherein i and k are any positive integer not greater than n, i is greater than k, and the resource category in the i-1 level is the superior resource category of the resource category in the ith level;

and determining the category characteristic corresponding to at least one resource category in the kth level as a first sample characteristic.

In some embodiments, determining a first resource category corresponding to the added behavior feature from among a plurality of resource categories includes:

acquiring newly added behavior data, wherein the newly added behavior data at least comprises at least one item of user data and at least one item of resource data, and the at least one item of resource data comprises a first resource type;

and extracting the characteristics of the data except the first resource type in the newly added behavior data to obtain the newly added behavior characteristics, and determining the first resource type as the first resource type corresponding to the newly added behavior characteristics.

In some embodiments, the merging the new behavior feature with the first class feature to obtain the updated first class feature of the first resource class includes:

and according to the weight of the newly added behavior feature and the first class feature, performing weighted fusion on the newly added behavior feature and the first class feature to obtain the first class feature after the first resource class is updated. .

In the embodiment of the disclosure, each newly added behavior feature is not stored separately, but the newly added behavior feature is fused with the corresponding stored category feature according to the resource category, so that the updated category feature fuses information included in the newly added behavior feature, the resource recommendation model can learn the newly added behavior feature when training according to the updated category feature, and the newly added behavior feature is not required to be stored again because the updated category feature is fused with the newly added behavior feature, so that the scheme improves the accuracy of training the resource recommendation model and reduces the storage pressure and the processing pressure.

FIG. 3 is a flow chart illustrating a method for resource recommendation, as shown in FIG. 3, including the following steps, according to an example embodiment.

301. And determining the resource characteristics of the resource to be recommended and the user characteristics of the target user account.

302. Determining model input characteristics based on the resource characteristics, the user characteristics and the category characteristics corresponding to a plurality of resource categories, wherein each category characteristic is obtained by respectively fusing the historical newly-added behavior characteristics corresponding to each resource category, and the historical newly-added behavior characteristics are determined according to historical behavior data.

303. And inputting the model input characteristics into the resource recommendation model to obtain the prediction recommendation parameters corresponding to the resources to be recommended.

304. And recommending the resources to be recommended to the target user account based on the predicted recommendation parameters.

In some embodiments, the method further comprises:

acquiring configuration information, wherein the configuration information comprises resource categories of n hierarchies, a plurality of resource categories are resource categories of the nth hierarchy in the n hierarchies, and n is a positive integer greater than 1;

determining model input features based on the resource features, the user features, and category features corresponding to the plurality of resource categories, including:

and determining the model input characteristics based on the resource characteristics, the user characteristics and the category characteristics corresponding to at least one resource category in the kth level, wherein k is a positive integer not larger than n.

In the embodiment of the disclosure, the resource recommendation model learns the interests of the user account in resources of different resource categories from a large number of newly added behavior features, so that whether the resource to be recommended is to be recommended to the target user account is determined through the resource recommendation model, and the accuracy of resource recommendation can be improved.

FIG. 4 is a flowchart illustrating another method for training a resource recommendation model according to an exemplary embodiment, where the execution subject is a server, as shown in FIG. 4, and the method includes the following steps:

401. the server determines a first resource category corresponding to the newly added behavior feature from a plurality of resource categories, wherein the newly added behavior feature is determined based on newly added behavior data, the newly added behavior data is generated based on the operation of a user account on the resource, and the first resource category is the category to which the resource belongs.

The newly added behavior characteristics can represent the behavior of the user on the resources, and can reflect the interest of the user on the resources. The resource is a multimedia resource, such as a video, audio, or news resource, or the resource is a commodity. The behavior is like, comment, collection, purchase, click or play. Optionally, the added behavior feature is represented in the form of an embedded (Embedding) vector. Different resources may belong to different categories; the classification mode of the resource and the number of resource categories may be set and changed as required, which is not limited by the present disclosure. For example, taking a resource as an example, the resource category can be classified into a game category, a music category, a movie category, a food category, or the like.

In the embodiment of the present disclosure, for each resource, the server determines the resource category to which the resource belongs first. In some embodiments, the server labels the resource category for each resource by way of manual labeling; or the server carries out Cluster analysis on a plurality of resources to obtain Cluster (Cluster) ID of each resource, and the Cluster ID is used as the resource category. Or, the server determines the resource category to which the resource belongs by adopting other implementation manners, which is not limited in the embodiment of the present disclosure.

In the embodiment of the disclosure, the number of the resource categories is flexibly set, so that the number of the resource categories can be elastically expanded and contracted according to different granularity requirements.

In some embodiments, the implementation of step 401 includes: the server acquires newly-added behavior data, wherein the newly-added behavior data at least comprises at least one item of user data and at least one item of resource data, and the at least one item of resource data comprises a first resource type; and extracting the characteristics of the data except the first resource type in the newly added behavior data to obtain the newly added behavior characteristics, and determining the first resource type as the first resource type corresponding to the newly added behavior characteristics.

The User data includes data such as a User (User) identifier, for example, the User identifier is an Identity Document (ID number). The resource data comprises data such as resource identification, author identification of resource author, resource category and the like. The resource author is a user account for uploading the resource. For example, taking a resource as a Video as an example, the resource data is Video data, the resource category is a Video category, and the Video data includes data such as a Video (Video) ID, an Author (Author) ID of a Video Author, and a Video category (Tag) ID.

In some embodiments, the new behavior data includes new sub-data with multiple dimensions, where the multiple dimensions include a user identifier dimension, a resource identifier dimension, or an author identifier dimension of a resource author, and accordingly, the new sub-data includes: user identification, resource identification, or author identification of the resource author, etc. When the characteristics of the newly added behavior data are extracted, the characteristics of the newly added subdata of each dimension are extracted respectively to obtain the characteristics of the newly added subdata of each dimension, and therefore the characteristics of the newly added behavior are formed by the obtained characteristics of the newly added subdata of multiple dimensions.

It should be noted that the new behavior data exists in the form of a behavior sequence, and therefore, the new behavior data further includes the time when the user performed the behavior on the resource. For example, the time is the time of the click, the time of the approval, and the like. In the case that the resource is a video or an audio, the newly added behavior data may further include a playing time length, where the playing time length is a time length for the terminal to play the resource.

Wherein the newly added behavior data is recorded for the terminal used by the user. Illustratively, each time a new piece of added behavior data is generated, the terminal sends the new piece of added behavior data to the server, and the server receives the new piece of added behavior data. Correspondingly, the server performs feature extraction on the newly added behavior data every time the server receives one piece of newly added behavior data, or the server stores the newly added behavior data first and performs feature extraction on a plurality of pieces of newly added behavior data respectively when receiving the plurality of pieces of newly added behavior data.

In the embodiment of the disclosure, the newly added behavior feature is obtained by performing feature extraction on the newly added behavior data, and the resource category corresponding to the newly added behavior feature is determined, so that data support is provided for subsequently updating the stored category feature according to the newly added behavior feature.

In some embodiments, the embodiment of the present disclosure only takes the execution subject as an example, and it should be noted that the server may include a processing server and a data storage server, where the processing server is configured to interact with each logged-in terminal, recommend a resource for a user account logged in by the terminal, or collect new behavior data generated by the user on the terminal, and the data storage server is configured to store data required by the recommendation server, such as the collected new behavior data, new behavior characteristics, or category characteristics stored in correspondence to a resource category. Then, the processing server collects the newly added behavior data, performs feature extraction on the newly added behavior data, stores the newly added behavior feature obtained after feature extraction in the data storage server, and the data storage server fuses the newly added behavior feature and the stored category feature to obtain an updated category feature for the processing server to use.

Illustratively, the data storage Server may be a PS (Parameter Server).

In the embodiment of the present disclosure, since the aggregation operation of the newly added behavior features is transferred from the processing server and the model layer to the data storage server and the Embedding vector layer, the method provided by the present disclosure does not need to Output more newly added behavior features from the data storage server to the processing server, but only needs to Output the category features obtained by fusion, thereby reducing the requirements on I/O (Input/Output) bandwidth and time delay between the processing server and the data storage server.

402. The server determines a first category characteristic corresponding to the first resource category.

Each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data. There are multiple resources in the server and different resources may have different resource categories. The server is divided according to resource categories and respectively stores the category characteristics corresponding to each resource category.

For example, for a plurality of added behavior features, if the added behavior features are directly stored, a large storage pressure is caused, and a large processing pressure is caused by subsequently reading the added behavior features. And after the newly added behavior features are fused to obtain category features, the category features are stored, the category features learn the information contained in the newly added behavior features, but the data volume is smaller than that of the newly added behavior features, so that the storage pressure is reduced, the reading of the category features is quicker and more efficient, and the processing pressure is also reduced. In some embodiments, the class feature is a Sparse (Sparse) feature fused from one or more added behavior features.

403. And the server fuses the newly added behavior characteristic and the first class characteristic to obtain the first class characteristic after the first resource class is updated.

The server fuses the newly added behavior features and the first class features, so that the first class features can learn the information contained in the newly added behavior features, and the updated first class features can represent the information contained in the newly added behavior features under the first resource class.

In some embodiments, taking an example that the server includes a processing server and a data storage server as an example for description, the processing server sends a fusion instruction to the data storage server, where the fusion instruction carries the newly added behavior feature, and the data storage server, in response to the fusion instruction, fuses the newly added behavior feature and the first class feature to obtain an updated first class feature, and stores the updated first class feature.

In some embodiments, the category characteristic corresponding to each resource category is stored in the bucket corresponding to the resource category, and accordingly, the implementation manner of step 403 includes: determining a first bucket corresponding to a first resource category from buckets corresponding to a plurality of resource categories; extracting a first class feature from the first bucket; and fusing the newly added behavior features and the first class features to obtain the updated first class features of the first resource class, and storing the updated first class features in the first bucket.

The server divides the class characteristics into buckets according to the resource classes, so that each resource class has a corresponding bucket, and the bucket stores the class characteristics corresponding to the resource class. In the embodiment of the present disclosure, the category characteristics are respectively stored in the form of buckets according to the resource categories, thereby facilitating the management of the category characteristics.

In a possible implementation manner of the foregoing embodiment, the new behavior feature includes new incremental sub-features of multiple dimensions, the new incremental sub-features of the multiple dimensions are obtained by feature extraction of new behavior data from different dimensions, and the new behavior data can be described from different dimensions. Each resource category corresponds to different buckets, each bucket corresponds to different dimensionality, and each bucket is used for storing a category sub-feature belonging to the corresponding dimensionality of the bucket under the resource category; and the category characteristics corresponding to each resource category comprise category sub-characteristics of a plurality of dimensions. Correspondingly, the server fuses the newly added behavior features with the first class features to obtain the updated first class features of the first resource class, and the implementation mode of storing the updated first class features in the first bucket includes: for each dimension in the plurality of dimensions, the server determines a second bucket corresponding to the dimension from the first bucket, and extracts the category sub-features from the second bucket; and fusing the new increasing sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in a second bucket.

For each resource category, the number of buckets corresponding to the resource category is the same as the number of dimensions corresponding to the resource category.

In some embodiments, the new added behavior data includes new added sub data of multiple dimensions, and the new added sub feature of each dimension included in the new added behavior feature is extracted based on the new added sub data corresponding to the dimension in the new added behavior data, for example, taking a resource as a video, and for a video ID dimension, the server performs feature extraction on a video ID in the new added behavior data to obtain a new added sub feature corresponding to the video ID.

For example, referring to fig. 5, the resource category is category a, the resource category is stored in correspondence with 3 buckets, the three buckets are bucket 1, bucket 2, and bucket 3, the corresponding dimension of each bucket is respectively a user identifier, a resource identifier, and an author identifier of a resource author, the server fuses the new adding sub-feature of the user identifier dimension in the new adding behavior feature and the category sub-feature stored in bucket 1, stores the updated category sub-feature in bucket 1, fuses the new adding sub-feature of the resource identifier dimension and the category sub-feature stored in bucket 2, stores the updated category sub-feature in bucket 2, fuses the new adding sub-feature of the author identifier dimension and the category sub-feature stored in bucket 3, and stores the updated category sub-feature in bucket 3.

In the embodiment of the disclosure, because the newly added behavior features and the category features both include features of multiple dimensions, by fusing the newly added sub-features and the category sub-features of each dimension, features of the same resource category and the same dimension can be fused together, so that there is no loss of information, and the utilization rate of the newly added behavior features is improved.

In some embodiments, the implementation of the server fusing the new behavior feature with the first class feature includes: and the server performs weighted fusion on the newly added behavior features and the first class features according to the weights of the newly added behavior features and the first class features to obtain the first class features after the first resource class is updated.

The weighted fusion may be based on a moving weighted average. For example, the sliding weighted average is an arithmetic sliding weighted average or an exponential sliding weighted average, which is not limited by this disclosure. Optionally, the server determines the weight of the newly added behavior feature and the first class feature as a fixed value, or the server determines the weight of the newly added behavior feature and the first class feature by combining with the reference score of the resource indicated by the newly added behavior feature. Wherein the reference score is a quality score or an importance score. A higher reference score indicates a higher quality or a higher importance of the resource. Correspondingly, the server sets higher weight for the newly added behavior characteristics with higher reference values. It should be noted that, the reference score of any resource may be set in advance according to needs, and the disclosure does not limit this.

Alternatively, the weighted fusion may be performed by an autoregressive-based sequence model. The autoregressive sequence model is an ARMA (Auto-Regressive and Moving Average) model, an LSTM (Long Short-Term Memory) model, or a transform model (an autoregressive model). In this embodiment, the newly added behavior feature and the first class feature are fused by the auto-regressive sequence model, so that a fusion result of the newly added behavior feature obtained from the history can be obtained.

In the embodiment of the disclosure, the newly added behavior features and the first class features are subjected to weighted fusion, so that the fused first class features, namely the updated first class features, can be combined with the information indicated by the historical newly added behavior features and the acquired information indicated by the newly added behavior features, compared with a mode of filtering part of the newly added behavior features to reduce storage pressure, the mode of storing the first class features does not cause information loss, and further data support is provided for training a more accurate resource recommendation model.

404. The server determines a first sample feature of the resource recommendation model based on category features corresponding to the multiple resource categories, wherein the category features corresponding to the multiple resource categories include the updated first category feature of the first resource category.

After the updated first class features are obtained, the server trains the resource recommendation model according to the class features corresponding to the plurality of currently stored resource classes. In some embodiments, the server directly splices category features corresponding to a plurality of resource categories to obtain a first sample feature of the resource recommendation model, and accordingly, the implementation manner of step 404 includes: for each resource category, the server respectively extracts category sub-features corresponding to the resource category from a plurality of buckets corresponding to the resource category; and splicing the extracted category sub-features corresponding to the plurality of resource categories to obtain a first sample feature.

The server splices the category sub-features corresponding to each resource category to obtain spliced category sub-features, and then splices the spliced category sub-features corresponding to the plurality of resource categories to obtain a first sample feature.

In the embodiment of the disclosure, the category sub-features stored in the buckets corresponding to the multiple resource categories are spliced to obtain a feature sequence, the feature sequence can represent information indicated by the historical newly added behavior features and information indicated by the obtained newly added behavior features, and the obtained data volume of the first sample feature is small, so that the input data volume of the resource recommendation model is reduced.

In some embodiments, the number of the plurality of resource categories may be greater, so that the server divides the resource categories of the plurality of hierarchies, and determines the first sample feature by combining the resource categories of the plurality of hierarchies, and then the method for training the resource recommendation model provided in the embodiment of the present disclosure further includes the following steps: the server obtains configuration information, wherein the configuration information comprises resource categories of n levels, the resource categories are resource categories of the nth level in the n levels, and n is a positive integer larger than 1. Accordingly, an implementation of step 404 includes:

the server fuses category characteristics corresponding to a plurality of resource categories of the same higher level in the ith level, determines the fused category characteristics as the category characteristics corresponding to the resource categories of the higher level, and determines the fused category characteristics as the category characteristics corresponding to at least one resource category in the kth level until the category characteristics corresponding to at least one resource category in the kth level are obtained, wherein i and k are any positive integer not greater than n, i is greater than k, and the resource category in the i-1 level is the higher level resource category of the resource category in the ith level; and determining the category characteristic corresponding to at least one resource category in the kth level as a first sample characteristic.

The number of stages of the kth stage may be set as required, which is not limited by the present disclosure. The resource categories in step 401 are resource categories of the nth level in the plurality of levels, that is, the server fuses the new adding behavior feature with the corresponding category feature according to the resource category of the last level each time.

Optionally, when the category features corresponding to a plurality of present-level resource categories belonging to the same upper-level resource category are fused, the server determines that the weight of each present-level resource category is a fixed numerical value, or determines the weight of each present-level resource category according to the reference score of each present-level resource category. For example, the reference score may be an importance score, and the server assigns higher weight to the level resource category with higher importance score. It should be noted that, the reference score of any resource category may be set in advance according to needs, and the disclosure does not limit this.

For example, the resource categories of the n hierarchies are a primary category, a secondary category and a tertiary category, respectively, where each primary category corresponds to at least one secondary category, each secondary category corresponds to at least one tertiary category, the number of the primary categories is 30, the number of the secondary categories is 200, and the number of the tertiary categories is 1000.

In the embodiment of the disclosure, resource categories of multiple hierarchies are flexibly divided, so that the fusion granularity of the elastic expansion newly-added behavior characteristics is supported, and further compromise between complexity and model effect can be achieved.

In some embodiments, after the category characteristics corresponding to a plurality of resource categories are obtained, the server recommends resources to the sample user account, and determines sample behavior data in combination with feedback of the sample user account on the resources, so that the resource recommendation model is trained in combination with the sample behavior data and the first sample characteristics. The sample user account is any user account, for example, the sample user account is a user account corresponding to the acquired new behavior characteristic or another user account.

Correspondingly, the training method of the resource recommendation model provided by the embodiment of the disclosure further includes the following steps (1) - (2):

(1) the server obtains sample behavior data, wherein the sample behavior data comprises resource characteristics, user characteristics and behavior labels of recommended sample resources.

The sample resources are recommended to the sample user account after category characteristics corresponding to a plurality of resource categories are obtained, the resource characteristics are used for describing the sample resources, the user characteristics are corresponding to the sample user account, and the behavior labels are used for representing types of behaviors of the sample user account on the sample resources.

The sample resource is a resource recommended to the sample user account after the category features of the stored multiple resource categories are updated according to the acquired new behavior features, that is, the currently acquired new behavior features do not include the resource features of the sample resource. The user characteristics may be user characteristics obtained by performing characteristic extraction on account information of the sample user account. The user characteristics may include characteristics corresponding to account data such as age data, gender data, or registration duration data of the sample user account. The resource characteristics can be characteristics corresponding to resource data, such as resource identification, author identification of resource author, browsing amount of the resource or applicable age of the resource.

For example, the sample resource is a video, and the resource feature is at least one of a feature corresponding to a video identifier, a feature corresponding to an author identifier of a video author, a feature corresponding to a browsing amount of the video, and the like. The type represented by the behavior tag comprises the types of behaviors such as praise, comment, collection, purchase, click, scratch, report or play. The behavior tag may also include other types, as the present disclosure is not limited in this respect.

(2) The server determines the resource characteristics and the user characteristics of the sample resources as second sample characteristics.

In the embodiment of the disclosure, on the basis of the first sample characteristic, sample behavior data is also obtained, so that a second sample characteristic is determined on the basis of the sample behavior data, and rich data support is provided for a subsequent training resource recommendation model.

405. The server trains the resource recommendation model based on the first sample characteristic, and the resource recommendation model is used for recommending resources to any user account.

The target user account is any user account of the resource to be recommended. In some embodiments, since the server further determines a second sample characteristic according to the obtained sample behavior data, the server trains the resource recommendation model in combination with the second sample characteristic on the basis of the first sample characteristic. Accordingly, the implementation of step 405 includes: the server trains a resource recommendation model based on the first sample features, the second sample features and the behavior labels. The behavior tag is a behavior tag included in the sample behavior data.

In the embodiment of the disclosure, in the training process of the resource recommendation model, reference is made to the historical behavior data corresponding to a plurality of resource categories, and reference is also made to the sample behavior data, so that the trained resource recommendation model has higher accuracy.

In some embodiments, the server trains an implementation of the resource recommendation model based on the first sample features, the second sample features, and the behavior tags, including the following steps (1) - (2):

(1) and the server fuses the first sample characteristic and the second sample characteristic to obtain a third sample characteristic.

Optionally, the server fuses the updated target features of the multiple resource categories and the resource features and the user features of the sample resources based on an attention mechanism. The server determines index information (query) based on the resource characteristics and the user characteristics of the sample resources, and uses the category characteristics corresponding to the multiple resource categories as key value information corresponding to the index information, so that the resource characteristics and the user characteristics of the sample resources are stored in correspondence with the category characteristics corresponding to the multiple resource categories, and subsequent retrieval is facilitated. The category feature can be regarded as an information storage, and the resource feature and the user feature of the sample resource are used as indexes, namely query vectors, and are used for searching and selecting part of information in the category feature.

The key (key) and the value (value) in the key value information are the same and are the category characteristics corresponding to a plurality of resource categories. By setting the keys and the values in the key value information as the updated category characteristics, the subsequent retrieval can not be confused due to different key values. For the determined index information, the server obtains key value information stored corresponding to the index information, namely category characteristics corresponding to a plurality of resource categories, determines similarity between the index information and the category characteristics corresponding to each resource category, and performs weighted average on the category characteristics corresponding to the plurality of resource categories based on the similarity to obtain a third sample characteristic.

For example, the Attention mechanism may be a Multi-head Attention (Multi-head Attention) mechanism or other Attention mechanism, which is not limited by this disclosure. The server splices the resource characteristics and the user characteristics of the sample resources to obtain index information; or the server takes the resource characteristics of the sample resources as index information; alternatively, the server uses the user characteristics as the index information, which is not limited in this disclosure.

It should be noted that, when the index information is the resource characteristics of the sample resources, the server fuses the resource characteristics of the sample resources and the category characteristics corresponding to the multiple resource categories, and splices the fused characteristics and the user characteristics to obtain a third sample characteristic; correspondingly, under the condition that the index information is the user characteristics of the sample resources, the server fuses the user characteristics of the sample resources and the category characteristics corresponding to the multiple resource categories, and splices the fused characteristics and the resource characteristics to obtain a third sample characteristic.

In the embodiment of the disclosure, the resource features and the user features of the sample resources are fused with the category features corresponding to a plurality of resource categories, so that the category features more critical to the currently recommended resources are focused on the category features among the numerous category features, the attention to other category features is reduced, and even irrelevant category features are filtered out, so that the problem of information overload can be solved, and the processing efficiency and the accuracy of a subsequent resource recommendation model are improved.

(2) And the server trains a resource recommendation model based on the third sample characteristics and the behavior labels.

The server determines a loss function of the resource recommendation model during training according to the third sample characteristic according to the behavior label of the sample resource, so that model parameters of the resource recommendation model are adjusted and optimized according to the loss function, and the resource recommendation model is updated and trained by combining the optimized model parameters.

In the embodiment of the disclosure, the resource recommendation model is trained based on the third sample characteristics and the behavior labels of the sample resources, so that the resource recommendation model can learn the interests of the sample user accounts in the resources of multiple resource categories, and the accuracy of the resource recommendation model training is improved.

In the embodiment of the disclosure, on one hand, the data size of the third sample features obtained by fusion is small, so that the data size of the training data input into the resource recommendation model is small, and the model training speed is high; on the other hand, the third sample characteristics not only contain information of the user account, information of resources recommended to the user account, but also contain information of newly-added behavior characteristics, so that the model can learn more characteristics during training, and the accuracy of training of the resource recommendation model is improved.

In other embodiments, the server, based on the first sample features, the second sample features, and the behavior labels, implements a method for training the resource recommendation model, comprising: and the server splices the first sample characteristic and the second sample characteristic to obtain a third sample characteristic, and trains a resource recommendation model based on the third sample characteristic and the behavior label.

The third sample feature obtained by splicing can completely reserve the category features corresponding to the multiple resource categories, the user features of the sample resources and the information indicated by the resource features. Optionally, the implementation manner of the server training the resource recommendation model based on the third sample feature and the behavior label is the same as the implementation manner of step (2) in step 405, and is not described herein again.

In the embodiment of the disclosure, the first sample characteristic, the second sample characteristic and the behavior label are directly spliced, so that the sample characteristic input into the resource recommendation model is more complete and abundant, and the accuracy of the resource recommendation model training can be improved.

For example, taking 100 resources as an example of the obtained new added behavior features, the new added behavior features include new added sub features whose dimensions are resource identifiers, the new added sub features are vectors with a length of 100, that is, the new added sub features include vectors with 100 resource identifiers, the number of resource categories is 30, the behavior tags are clicks, the new added sub features are fused according to the resource categories and are fused into 30 buckets with different resource categories, and the category sub features updated in the 30 buckets can form category features with a length of 30.

In the embodiment of the present disclosure, the stored category features are obtained based on fusion of all the behavior features, and there is no loss of features because there is no eviction and selection of behavior features during fusion. The more information that is contained in the category features to represent the behavior of the user account on the resource as the resource recommendation model is trained. When the resource recommendation model is trained for a sufficiently long time, it can be considered that these class features utilize infinitely long behavior features.

In the embodiment of the disclosure, since the stored class characteristics are fused only based on the newly added behavior characteristics, that is, the newly generated behavior characteristics, the full amount of behavior characteristics do not need to be stored and read when the resource recommendation model is trained; and the behavior characteristics are not selected, but the resource recommendation model is updated according to the category characteristics updated each time, so that the requirements on the storage capacity and the throughput capacity of the server are reduced, and the accuracy of the resource recommendation model training is improved.

In the embodiment of the disclosure, each newly added behavior feature is not stored separately, but the newly added behavior feature is fused with the corresponding stored category feature according to the resource category, so that the updated category feature fuses information included in the newly added behavior feature, the resource recommendation model can learn the newly added behavior feature when training according to the updated category feature, and the newly added behavior feature is not required to be stored again because the updated category feature is fused with the newly added behavior feature, so that the scheme improves the accuracy of training the resource recommendation model and reduces the storage pressure and the processing pressure.

FIG. 6 is a flowchart illustrating another method for training a resource recommendation model according to an exemplary embodiment, where the execution subject is a server, as shown in FIG. 6, and the method includes the following steps:

601. the server determines a plurality of first resource categories corresponding to the newly added behavior characteristics from a plurality of resource categories, wherein the newly added behavior characteristics are determined based on newly added behavior data, the newly added behavior data are generated based on the operation of the user account on the resource, and each first resource category is the category to which the resource belongs.

In some embodiments, the implementation manner of step 601 is the same as that of step 401, and is not described herein again. Optionally, the same resource may belong to multiple resource categories, in this embodiment of the disclosure, taking the resource corresponding to multiple first resource categories as an example, correspondingly, the server determines the first category feature corresponding to each first resource category, respectively. For example, taking a resource as an example of a video, the plurality of first resource categories to which the video belongs include a game category and a music category.

602. The server determines first category features stored respectively corresponding to each of the first resource categories.

In some embodiments, the implementation manner of step 602 is the same as that of step 402, and is not described herein again.

603. For each first resource type, the server fuses the newly added behavior feature and the first type feature corresponding to the first resource to obtain the updated first type feature of the first resource type.

In some embodiments, the implementation of the server fusing the new behavior feature with the first class feature includes: the server determines the newly added behavior characteristics distributed for each first resource type according to the product of the distribution proportion corresponding to each first resource type and the newly added behavior characteristics; and for each first resource category, fusing the newly added behavior characteristics obtained by the allocation of the first resource category with the first category characteristics corresponding to the first resource category to obtain the first category characteristics after the updating of the first resource category.

The distribution ratio may be set as needed, which is not limited by the present disclosure. It should be noted that, because the new behavior feature includes new increment sub-features of multiple dimensions, when the new behavior feature is allocated, the new increment sub-features of each dimension are allocated according to the allocation proportion.

In this embodiment, for each first resource category, the implementation manner of the server fusing the added behavior feature allocated by the first resource category with the first category feature corresponding to the first resource category is the same as the implementation manner of the server fusing the added behavior feature with the first category feature in step 403, and details are not described here again.

In the embodiment of the disclosure, when the resource indicated by the newly added behavior feature belongs to a plurality of first resource categories, the updated first category feature is more accurate by performing proportional allocation on the newly added behavior feature, so that the accuracy of the resource recommendation model training can be improved.

For example, with continued reference to fig. 5, for each first resource category, the server fuses the assigned new sub-features of different dimensions in the new behavior features with the extracted category sub-features of the dimension, and stores the fused category sub-features in the bucket corresponding to the dimension.

604. The server determines first sample features of the resource recommendation model based on category features corresponding to the plurality of resource categories, wherein the category features corresponding to the plurality of resource categories include the updated first category features of each first resource category.

605. The server trains the resource recommendation model based on the first sample characteristic, and the resource recommendation model is used for recommending resources to any user account.

In some embodiments, the implementation manner of step 604-605 is the same as that of step 404-405, and will not be described herein again.

In the embodiment of the disclosure, each newly added behavior feature is not stored separately, but the newly added behavior feature is fused with the corresponding stored category feature according to the resource category, so that the updated category feature fuses information included in the newly added behavior feature, the resource recommendation model can learn the newly added behavior feature when training according to the updated category feature, and the newly added behavior feature is not required to be stored again because the updated category feature is fused with the newly added behavior feature, so that the scheme improves the accuracy of training the resource recommendation model and reduces the storage pressure and the processing pressure.

In the embodiment of the present disclosure, since the trained resource recommendation model already learns a large number of behavior features, the accuracy of resource recommendation using the resource recommendation model is high, so that a server can perform resource recommendation by using the resource recommendation model, and accordingly, the resource recommendation to a target user account is described as an example, fig. 7 is a flowchart of a resource recommendation method according to an exemplary embodiment, and as shown in fig. 7, an execution subject is a server, and the method includes the following steps:

701. the server determines the resource characteristics of the resource to be recommended and the user characteristics of the target user account.

For the resource to be recommended, the server extracts the characteristics of the resource to obtain the characteristics of the resource; and the server performs characteristic extraction on the account of the target user account to obtain the user characteristics.

702. The server determines model input features based on the resource features, the user features and category features corresponding to a plurality of resource categories, wherein each category feature is obtained by respectively fusing historical behavior features corresponding to each resource category, and the historical behavior features are determined according to historical behavior data.

In some embodiments, since hierarchical relationships may exist between resource categories, when acquiring category features currently corresponding to a plurality of resource categories, the server acquires a category feature corresponding to a resource category of any one of the plurality of hierarchies; correspondingly, the resource recommendation method provided by the embodiment of the disclosure further includes the following steps: the server obtains configuration information, wherein the configuration information comprises resource categories of n levels, a plurality of resource categories are resource categories of the nth level in the n levels, and n is a positive integer greater than 1. Accordingly, the implementation of step 702 includes: the server determines the model input characteristics based on the resource characteristics, the user characteristics and the category characteristics corresponding to at least one resource category in the kth level, wherein k is a positive integer not greater than n.

The implementation manner of the category feature corresponding to the at least one resource category in the kth level determined by the server is the same as the implementation manner of the category feature corresponding to the at least one resource category in the kth level determined by the server in step 404, and details are not repeated here.

In the embodiment of the disclosure, the category characteristics currently corresponding to a plurality of resource categories can be determined in a plurality of hierarchies, so that the data size of the model input characteristics can be flexibly set according to the needs of the model, and the adaptability of the resource recommendation model is improved.

703. And the server inputs the model input characteristics into the resource recommendation model to obtain the prediction recommendation parameters corresponding to the resources to be recommended.

The training process of the resource recommendation model is as described in the above embodiments. The server takes the model input characteristics as input data of a resource recommendation model, and determines a prediction recommendation parameter of the resource to be recommended through the resource recommendation model, wherein the prediction recommendation parameter represents the acceptance degree of the target user account on the resource to be recommended.

704. And the server recommends the resources to be recommended to the target user account based on the predicted recommendation parameters.

In some embodiments, the predicted recommendation parameters meet the recommendation conditions and indicate that the acceptance degree of the target user account for the resources to be recommended is higher, and the server recommends the resources to be recommended to the target user account; and if the predicted recommendation parameters do not meet the recommendation conditions and indicate that the target user account has low acceptance degree of the resources to be recommended, the server does not need to recommend the resources to be recommended to the target user account.

Optionally, the prediction recommendation parameter is represented in a form of a score, and the recommendation condition is higher than the recommendation score, and the recommendation score may be set as needed, which is not limited by the present disclosure. For example, the recommendation score is 80, 90, or 95, etc.

In the embodiment of the disclosure, the resource recommendation model learns the interests of the user account in resources of different resource categories from a large number of newly added behavior characteristics, so that the resource recommendation model is used to determine whether to recommend the resource to be recommended to the target user account, so that the resource recommendation accuracy can be improved.

FIG. 8 is a block diagram illustrating an architecture of a training apparatus for a resource recommendation model according to an exemplary embodiment. Referring to fig. 8, the apparatus includes a first determination unit 801, a fusion unit 802, and a training unit 803:

a first determining unit 801 configured to perform determining, from a plurality of resource categories, a first resource category corresponding to a newly added behavior feature, the newly added behavior feature being determined based on newly added behavior data, the newly added behavior data being generated based on an operation of a resource by a user account, the first resource category being a category to which the resource belongs; each resource category corresponds to different category characteristics, each category characteristic is obtained by respectively fusing historical behavior characteristics corresponding to each resource category, and the historical behavior characteristics are determined according to historical behavior data;

the fusion unit 802 is configured to perform determining a first class feature corresponding to the first resource class, and fuse the newly added behavior feature with the first class feature to obtain an updated first class feature of the first resource class;

the training unit 803 is configured to execute determining a first sample feature of a resource recommendation model based on category features corresponding to multiple resource categories, where the category features corresponding to the multiple resource categories include the updated first category feature of the first resource category, and train the resource recommendation model based on the first sample feature, where the resource recommendation model is used to recommend resources to any user account.

In some embodiments, referring to fig. 9, the class feature corresponding to each resource class is stored in a bucket corresponding to the resource class, and the merging unit 802 includes:

a determining subunit 8021, configured to perform determining, from buckets corresponding to multiple resource categories, a first bucket corresponding to a first resource category;

an extraction subunit 8022 configured to perform extraction of the first class feature from the first bucket;

the merging subunit 8023 is configured to perform merging the new added behavior feature with the first category feature, to obtain an updated first category feature of the first resource category, and store the updated first category feature in the first bucket.

In some embodiments, the added behavior feature comprises a plurality of dimensions of added sub-features;

each resource category corresponds to different buckets, each bucket corresponds to different dimensionality, and each bucket is used for storing a category sub-feature belonging to the corresponding dimensionality of the bucket under the resource category; the category characteristics corresponding to each resource category comprise category sub-characteristics of a plurality of dimensions;

a fusion subunit 8023 configured to perform, for each dimension of the plurality of dimensions, determining a second bucket corresponding to the dimension from the first bucket, extracting a category sub-feature from the second bucket; and fusing the new increasing sub-features of the dimensionality and the extracted category sub-features, and storing the fused category sub-features in a second bucket.

In some embodiments, the training unit 803 is configured to perform, for each resource category, extracting category sub-features corresponding to the resource category from a plurality of buckets corresponding to the resource category, respectively; and splicing the extracted category sub-features corresponding to the plurality of resource categories to obtain a first sample feature.

In some embodiments, the apparatus further comprises:

a first obtaining unit 804, configured to perform obtaining of sample behavior data, where the sample behavior data includes resource features, user features, and behavior tags of recommended sample resources, where the sample resources are resources recommended to a sample user account after category features corresponding to multiple resource categories are obtained, the resource features are used to describe the sample resources, the user features are user features corresponding to the sample user account, and the behavior tags are used to indicate types of behaviors performed on the sample resources by the sample user account;

a second determining unit 805 configured to perform determining the resource feature and the user feature of the sample resource as a second sample feature;

a training unit 803 configured to perform training of the resource recommendation model based on the first sample features, the second sample features and the behavior labels.

In some embodiments, the training unit 803 is configured to perform stitching of the first sample feature and the second sample feature to obtain a third sample feature, and train the resource recommendation model based on the third sample feature and the behavior label; alternatively, the first and second electrodes may be,

a training unit 803, configured to perform fusion of the first sample feature and the second sample feature to obtain a third sample feature, and train the resource recommendation model based on the third sample feature and the behavior label.

In some embodiments, the resource belongs to a plurality of first resource categories, and the merging unit 802 is configured to perform a multiplication of an allocation proportion corresponding to each first resource category and a new-added behavior feature, and determine a new-added behavior feature allocated to each first resource category; and for each first resource category, fusing the newly added behavior characteristics obtained by the allocation of the first resource category with the first category characteristics corresponding to the first resource category to obtain the first category characteristics after the updating of the first resource category.

In some embodiments, the apparatus further comprises:

a second obtaining unit 806, configured to perform obtaining configuration information, where the configuration information includes resource categories of n hierarchies, where a plurality of resource categories are resource categories of an nth level in the n hierarchies, and n is a positive integer greater than 1;

a training unit 803 configured to perform fusion of category features corresponding to a plurality of present-level resource categories belonging to the same superior resource category in an i-th level, and determine the fused category features as category features corresponding to a superior resource category until a category feature corresponding to at least one resource category in a kth level is obtained, where i and k are any positive integers not greater than n, and i is greater than k, where the resource category in the i-1 th level is the superior resource category of the resource category in the i-th level; and determining the category characteristic corresponding to at least one resource category in the kth level as a first sample characteristic.

In some embodiments, the first determining unit 801 is configured to perform obtaining new added behavior data, where the new added behavior data includes at least one item of user data and at least one item of resource data, and the at least one item of resource data includes a first resource category; and extracting the characteristics of the data except the first resource type in the newly added behavior data to obtain the newly added behavior characteristics, and determining the first resource type as the first resource type corresponding to the newly added behavior characteristics.

In some embodiments, the fusion unit 802 is configured to perform weighted fusion on the newly added behavior feature and the first class feature according to the weight of the newly added behavior feature and the first class feature, so as to obtain the updated first class feature of the first resource class.

In the embodiment of the disclosure, each newly added behavior feature is not stored separately, but the newly added behavior feature is fused with the corresponding stored category feature according to the resource category, so that the updated category feature fuses information included in the newly added behavior feature, the resource recommendation model can learn the newly added behavior feature when training according to the updated category feature, and the newly added behavior feature is not required to be stored again because the updated category feature is fused with the newly added behavior feature, so that the scheme improves the accuracy of training the resource recommendation model and reduces the storage pressure and the processing pressure.

Fig. 10 is a block diagram illustrating a configuration of a resource recommendation apparatus according to an exemplary embodiment. Referring to fig. 10, the apparatus includes a third determining unit 1001, a fourth determining unit 1002, a fifth determining unit 1003, and a recommending unit 1004:

a third determining unit 1001 configured to perform determining a resource characteristic of a resource to be recommended and a user characteristic of a target user account;

a fourth determining unit 1002, configured to perform determining a model input feature based on the resource feature, the user feature, and category features currently corresponding to multiple resource categories, where each category feature is obtained by respectively fusing historical behavior features corresponding to each resource category, and the historical behavior features are determined according to historical behavior data;

a fifth determining unit 1003 configured to input the model input characteristics into the resource recommendation model, so as to obtain a predicted recommendation parameter corresponding to the resource to be recommended;

and a recommending unit 1004 configured to recommend the resource to be recommended to the target user account based on the predicted recommending parameter.

In some embodiments, referring to fig. 11, the apparatus further comprises:

a third obtaining unit 1005 configured to perform obtaining configuration information, the configuration information including resource categories of n hierarchies, where the plurality of resource categories are resource categories of an nth level in the n hierarchies, and n is a positive integer greater than 1;

a fourth determining unit 1002 configured to perform determining the model input feature based on the resource feature, the user feature and the category feature corresponding to at least one resource category in the kth level, k being a positive integer not greater than n.

In the embodiment of the disclosure, the resource recommendation model learns the interests of the user account in resources of different resource categories from a large number of behavior characteristics, so that whether the resource to be recommended is to be recommended to the target user account is determined by the resource recommendation model, and the accuracy of resource recommendation can be improved.

With regard to the training apparatus of the resource recommendation model and the resource recommendation apparatus in the above embodiments, the specific manner in which each unit performs the operation has been described in detail in the embodiments of the related method, and will not be elaborated herein.

Fig. 12 is a block diagram illustrating a server 1200, which may have a relatively large difference due to different configurations or performances according to an exemplary embodiment, and may include one or more processors (CPUs) 1201 and one or more memories 1202 for storing executable instructions of the processors, where the processors 1201 are configured to execute the instructions to implement the training method or the resource recommendation method of the resource recommendation model in the above embodiments. Certainly, the server 1200 may further have components such as a wired or wireless network interface, a keyboard, and an input/output interface, so as to perform input and output, and the server 1200 may further include other components for implementing the functions of the device, which is not described herein again.

In an exemplary embodiment, there is also provided a computer-readable storage medium including instructions, which when executed by a processor of a server, enable the server to perform the training method of the resource recommendation model or the resource recommendation method in the above embodiments. For example, the computer-readable storage medium may be a ROM (Read Only Memory), a RAM (Random Access Memory), a CD-ROM (Compact Disc Read-Only Memory), a magnetic tape, a floppy disk, an optical data storage device, and the like.

In an exemplary embodiment, a computer program product is also provided, which comprises a computer program, which when executed by a processor implements the method for training a resource recommendation model or the method for resource recommendation in the above-mentioned embodiments.

In some embodiments, the computer program according to the embodiments of the present disclosure may be deployed to be executed on one server or on a plurality of servers located at one site, or may be executed on a plurality of servers distributed at a plurality of sites and interconnected by a communication network, and the plurality of servers distributed at the plurality of sites and interconnected by the communication network may constitute a block chain system.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.