阅读说明：本技术 一种用于去除系统推荐偏差的方法和系统 (Method and system for removing system recommendation deviation ) 是由 周晔 穆海洁 景晓峰 于 2021-09-29 设计创作，主要内容包括：本发明公开了一种用于去除系统推荐偏差的方法和系统,能够用一个通用的去偏差框架解决各种情况下偏差对推荐系统的影响。其技术方案为：接收推荐系统场景中与推荐偏差相关的输入数据,作为数据特征层的一个或多个特征组；将接收到的数据按照特征组的类别进行向量处理,将数据的特征值转化为向量值,以将特征组分别转化为对应的向量组,再将一个或多个向量组传递至多层神经网络,在多层神经网络中将向量组中的词向量交叉组合成数据集,以这些数据集构建出共享数据层；将共享数据层中的数据划分成多个不同的数据组,并对该多个数据组采用对应的去偏策略进行去偏处理,将所有数据组经去偏处理后的数据整合输出。(The invention discloses a method and a system for removing system recommendation deviation, which can solve the influence of deviation on a recommendation system under various conditions by using a universal deviation removing framework. The technical scheme is as follows: receiving input data related to recommendation deviation in a recommendation system scene as one or more feature groups of a data feature layer; carrying out vector processing on received data according to the categories of the feature groups, converting feature values of the data into vector values so as to convert the feature groups into corresponding vector groups respectively, transmitting one or more vector groups to a multilayer neural network, and cross-combining word vectors in the vector groups into data sets in the multilayer neural network so as to construct a shared data layer by the data sets; dividing the data in the shared data layer into a plurality of different data groups, carrying out depolarization processing on the plurality of data groups by adopting corresponding depolarization strategies, and integrating and outputting the data after the depolarization processing of all the data groups.)

1. A method for removing system recommendation bias, the method comprising:

step 1: receiving input data related to recommendation deviation in a recommendation system scene as one or more feature groups of a data feature layer;

step 2: carrying out vector processing on received data according to the categories of the feature groups, converting feature values of the data into vector values so as to convert the feature groups into corresponding vector groups respectively, transmitting one or more vector groups to a multilayer neural network, and cross-combining word vectors in the vector groups into data sets in the multilayer neural network so as to construct a shared data layer by the data sets;

and step 3: dividing the data in the shared data layer into a plurality of different data groups, carrying out depolarization processing on the plurality of data groups by adopting corresponding depolarization strategies, and integrating and outputting the data after the depolarization processing of all the data groups.

2. The method for removing system recommendation deviation according to claim 1, wherein the step 2 is to convert the characteristic value into a vector value through an embedding process in a neural network.

3. The method for removing the system recommendation deviation according to claim 1, wherein the deviation removing strategy in step 3 comprises: a trend score method, a data filling method, a dual robust estimator and joint learning.

4. The method for removing systematic recommendation bias according to claim 1, wherein the partitioning of the data group in step 3 is based on different task objectives.

5. A system for removing system recommended deviation is characterized in that the system comprises a data preprocessing module, a deviation removal model module and a multi-target model module, wherein the data preprocessing module comprises a characteristic value input unit, a characteristic value vectorization unit and a multi-layer neural network processing unit, the deviation removal model module is composed of one or more deviation processing units, the multi-target model module is composed of one or more deviation target grouping units, and the method comprises the following steps:

the characteristic value input unit is used for receiving input data related to recommendation deviation in a recommendation system scene as one or more characteristic groups of a data characteristic layer;

the characteristic value vectorization unit is used for carrying out vector processing on the received data according to the types of the characteristic groups, converting the characteristic values of the data into vector values so as to respectively convert the characteristic groups into corresponding vector groups, and then transmitting one or more vector groups to the multilayer neural network;

the multilayer neural network processing unit is used for crossly combining word vectors in the vector group into data sets in the multilayer neural network, and constructing a shared data layer by using the data sets;

the multi-target model module is used for dividing data in a shared data layer constructed by the multi-layer neural network processing unit into a plurality of different data groups through a strategy;

and the depolarization model module is used for carrying out depolarization processing on the plurality of data groups divided by the multi-target model module by adopting a corresponding depolarization strategy, and integrating and outputting data after the depolarization processing of all the data groups.

6. The system for removing system recommendation deviation according to claim 5, wherein the multi-objective model module comprises: a position deviation target grouping unit, a selection deviation target grouping unit and an exposure deviation grouping unit; the depolarization model module comprises: a position deviation processing unit, a selection deviation processing unit and an exposure deviation processing unit.

7. The system for removing system recommendation deviation according to claim 5, wherein the feature value vectorization unit converts the feature value into a vector value through an embedding process in a neural network.

8. The system for removing system recommendation deviation according to claim 5, wherein the partitioning of the data groups in the multi-objective model module is based on different task objectives.

9. A computer system for removing system recommended deviations, the computer system comprising:

a processor;

a memory configured to store a series of computer-executable instructions and computer-accessible data associated with the series of computer-executable instructions;

the series of computer executable instructions, when executed by the processor, cause the processor to perform the method for removing system recommendation bias of any of claims 1 to 4.

10. A non-transitory computer readable storage medium having stored thereon a series of computer executable instructions which, when executed by a computing device, cause the computing device to perform the method for removing system recommendation bias of any of claims 1 to 4.

Technical Field

The invention relates to a technology for removing system recommendation deviation, in particular to a method and a system for removing the system recommendation deviation by adopting a multi-target model.

Background

The recommendation system is taken as an important means for solving information overload and providing personalized content for users, and remarkable results are obtained in application scenes of various industries nowadays. The traditional recommendation system collects user behavior data as original data, combines technologies such as machine learning and the like, and finally returns recommended content to the user. However, various deviations often exist in the process of collecting user behavior data, so that the recommendation effect is reduced.

The common deviations are mainly concentrated as follows:

selecting deviation: users prefer to score favorite or offensive items, resulting in items between the two emotions lacking scoring data;

position deviation: the influence of the sorting position factor on the clicking behavior of the user, and the user tends to click more advanced contents under the normal condition;

exposure deviation: due to the influence of hot articles or previous recommendation results, a user can receive certain articles more easily, and other articles cannot be exposed correspondingly, so that articles which are interested by the user are not displayed to the user, and deviation is brought.

The existing technology for removing common deviations is usually aimed at one or two specific deviations, but in a real application scene, multiple deviations are often mixed together and cross-affected. Therefore, a general de-biasing framework is needed for the recommendation system to solve the influence of the bias on the recommendation system in various situations.

Disclosure of Invention

The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

The invention aims to solve the problems and provides a method and a system for removing system recommendation deviation, which can solve the influence of deviation on a recommendation system in various situations by using a universal deviation removing framework.

The technical scheme of the invention is as follows: the invention discloses a method for removing system recommendation deviation, which comprises the following steps:

step 1: receiving input data related to recommendation deviation in a recommendation system scene as one or more feature groups of a data feature layer;

step 2: carrying out vector processing on received data according to the categories of the feature groups, converting feature values of the data into vector values so as to convert the feature groups into corresponding vector groups respectively, transmitting one or more vector groups to a multilayer neural network, and cross-combining word vectors in the vector groups into data sets in the multilayer neural network so as to construct a shared data layer by the data sets;

and step 3: dividing the data in the shared data layer into a plurality of different data groups, carrying out depolarization processing on the plurality of data groups by adopting corresponding depolarization strategies, and integrating and outputting the data after the depolarization processing of all the data groups.

According to an embodiment of the method for removing the systematic recommendation deviation, in step 2, the feature values are converted into vector values through an embedding process in a neural network.

According to an embodiment of the method for removing the system recommendation deviation of the present invention, the deviation removing strategy in step 3 includes: a trend score method, a data filling method, a dual robust estimator and joint learning.

According to an embodiment of the method for removing systematic recommendation bias of the present invention, the division of the data group in step 3 is divided according to different task objectives.

The invention also discloses a system for removing the system recommendation deviation, which comprises a data preprocessing module, a deviation removal model module and a multi-target model module, wherein the data preprocessing module comprises a characteristic value input unit, a characteristic value vectorization unit and a multi-layer neural network processing unit, the deviation removal model module consists of one or more deviation processing units, the multi-target model module consists of one or more deviation target grouping units, and the method comprises the following steps:

the characteristic value input unit is used for receiving input data related to recommendation deviation in a recommendation system scene as one or more characteristic groups of a data characteristic layer;

the characteristic value vectorization unit is used for carrying out vector processing on the received data according to the types of the characteristic groups, converting the characteristic values of the data into vector values so as to respectively convert the characteristic groups into corresponding vector groups, and then transmitting one or more vector groups to the multilayer neural network;

the multilayer neural network processing unit is used for crossly combining word vectors in the vector group into data sets in the multilayer neural network, and constructing a shared data layer by using the data sets;

the multi-target model module is used for dividing data in a shared data layer constructed by the multi-layer neural network processing unit into a plurality of different data groups through a strategy;

and the depolarization model module is used for carrying out depolarization processing on the plurality of data groups divided by the multi-target model module by adopting a corresponding depolarization strategy, and integrating and outputting data after the depolarization processing of all the data groups.

According to an embodiment of the system for removing system recommendation deviation of the present invention, the multi-target model module comprises: a position deviation target grouping unit, a selection deviation target grouping unit and an exposure deviation grouping unit; the depolarization model module comprises: a position deviation processing unit, a selection deviation processing unit and an exposure deviation processing unit.

According to an embodiment of the system for removing the system recommendation deviation, the feature value is converted into a vector value through an embedding process in a neural network in the feature value vectorization unit.

According to an embodiment of the system for removing system recommendation deviation of the present invention, the division of the data group in the multi-target model module is divided according to different task objectives.

The invention also discloses a computer system for removing the systematic recommended deviation, which comprises:

a processor;

a memory configured to store a series of computer-executable instructions and computer-accessible data associated with the series of computer-executable instructions;

the series of computer executable instructions, when executed by the processor, cause the processor to perform the method for removing system recommendation bias as described above.

Also disclosed is a non-transitory computer readable storage medium having stored thereon a series of computer executable instructions which, when executed by a computing device, cause the computing device to perform the method for removing system recommendation bias as described above.

Compared with the prior art, the invention has the following beneficial effects: the traditional unbiasing method is often to bring in the bias feature when the feature is input, so as to optimize the input of the feature value.

The invention has the innovation points that the depolarization processing process is arranged in the multi-target model processing stage, the multi-target model and the depolarization model have the same action, and different deviation strategies are used under different target scenes while the data processing efficiency is improved through the multi-target model, so that the depolarization effect is achieved. The multi-target model can be adjusted in a self-defined mode, for example, under the scene that specific single deviation needs to be processed, the corresponding target task model is adjusted to be larger than the target task model, and therefore the specific deviation condition is processed; under the general situation, the deviation of the general recommendation system can be directly removed without adjusting the multi-target model.

Drawings

The above features and advantages of the present disclosure will be better understood upon reading the detailed description of embodiments of the disclosure in conjunction with the following drawings. In the drawings, components are not necessarily drawn to scale, and components having similar relative characteristics or features may have the same or similar reference numerals.

FIG. 1 illustrates a flow diagram of one embodiment of a method of the present invention for removing systematic recommendation biases.

FIG. 2 illustrates a schematic diagram of an embodiment of a system for removing systematic recommendation biases of the present invention.

FIG. 3 illustrates a schematic diagram of one embodiment of a computer system for removing system recommendation biases in accordance with the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. It is noted that the aspects described below in connection with the figures and the specific embodiments are only exemplary and should not be construed as imposing any limitation on the scope of the present invention.

FIG. 1 illustrates a flow diagram of one embodiment of a method of the present invention for removing systematic recommendation biases. Referring to fig. 1, the specific implementation steps of the method of the present embodiment are detailed as follows.

Step 1: input data related to the recommendation deviation in the recommendation system scenario is received as one or more feature groups of the data feature layer.

The input data related to the recommendation deviation received in step 1 may vary according to the specific scenario of the recommendation system. For example, in a video recommendation scene, the feature groups may be a user historical behavior feature group (like praise, collection, viewing duration, etc.), a user context feature group (user device information, viewing time, etc.), a video feature group (video type, video tag, etc.), etc., and therefore the types and the numbers of the feature groups may be different according to different service scenes.

Step 2: the method comprises the steps of carrying out vector processing on received data according to the categories of feature groups, converting feature values of the data into vector values to convert the feature groups into corresponding vector groups respectively, transmitting one or more vector groups to a multilayer neural network, and cross-combining word vectors in the vector groups into data sets in the multilayer neural network to construct a shared data layer by the data sets.

The vector processing for converting the characteristic value into the vector value is an embedded technology, called Embedding, and is a technology for converting a discrete variable into a continuous vector in the field of neural networks.

And step 3: dividing the data in the shared data layer into a plurality of different data groups, carrying out depolarization processing on the plurality of data groups by adopting corresponding depolarization strategies, and integrating and outputting the data after the depolarization processing of all the data groups.

In step 3, because different deskewing strategies are applied to different data sets, this embodiment essentially constructs a universal deskewing model by combining a plurality of different deskewing strategies, classifies data processed by the neural network according to the multipath deskewing strategy shown in fig. 1, then respectively removes the biases of the multipath data, and finally outputs the result to the output layer.

In order to ensure that the effect of the division of the data group (namely the division of the target task) is better, after the processing of the multilayer neural network layer in the preorder step, the input data is subjected to multilayer cross processing, the data dimensionality of each piece of data in the shared data layer tends to be consistent, and at the moment, the data only needs to be classified according to the target requirement; the specific data scale division can also be performed by combining the scale of the input data in the step 1: the more data is divided by the multi-objective model, which type of data is input more.

The depolarization strategy comprises the following steps: a trend score method, a data filling method, a dual robust estimator, joint learning, etc.

Trend scoring refers to a class of statistical methods that use non-experimental or observed data for intervention effect analysis. The data filling method includes a mean value filling method and the like. A dual robust estimator refers to the use of an adaptive filter to select targets from clutter (other signals than the desired signal). Joint learning refers to the client training the model together under the coordination of the central server, while keeping decentralization and dispersion of the training data.

Dividing the data group is equivalent to dividing a plurality of target tasks based on different task purposes, different depolarization strategies are adopted for different data groups, and corresponding depolarization processing is equivalent to using the corresponding depolarization strategies for different task targets. The division of the data groups and the selection of the corresponding deskewing strategy based on each group of data are divided according to different task purposes, for example, a target task for eliminating the position deviation is established currently, and then the recalled data can be recalled according to the characteristic value of the position parameter of the input data by using a tendency score method when the input data is referred to.

FIG. 2 illustrates the principles of one embodiment of the system for removing systematic recommendation biases of the present invention. Referring to fig. 2, the system includes a data preprocessing module, a depolarization model module, and a multi-objective model module.

The data preprocessing module comprises a characteristic value input unit, a characteristic value vectorization unit and a multilayer neural network processing unit.

The depolarization model module is composed of one or more deviation processing units, such as a position deviation processing unit, a selection deviation processing unit, and an exposure deviation processing unit.

The multi-target model module is composed of one or more deviation target grouping units, such as a position deviation target grouping unit, a selection deviation target grouping unit and an exposure deviation grouping unit.

Each deviation target grouping unit in the multi-target model module corresponds to a deviation processing unit in the deviation removal model module.

The characteristic value input unit transmits the characteristic value of the input data to the characteristic value vectorization unit; the eigenvalue vectorization unit transmits the eigenvalue vector to the multilayer neural network processing unit; the multi-layer neural network processing unit transmits the characteristic data to the multi-target model module; the data preprocessing module transmits the characteristic data processed by the neural network to the multi-target model module, and the multi-target model and the depolarization model act together to achieve the depolarization effect.

The characteristic value input unit is used for receiving input data related to recommendation deviation in a recommendation system scene as one or more characteristic groups of a data characteristic layer.

The input data related to the recommendation deviation received by the characteristic value input unit can be different according to different specific scenes of the recommendation system. For example, in a video recommendation scene, the feature groups may be a user historical behavior feature group (like praise, collection, viewing duration, etc.), a user context feature group (user device information, viewing time, etc.), a video feature group (video type, video tag, etc.), etc., and therefore the types and the numbers of the feature groups may be different according to different service scenes.

The characteristic value vectorization unit is used for carrying out vector processing on the received data according to the types of the characteristic groups, converting the characteristic values of the data into vector values so as to respectively convert the characteristic groups into corresponding vector groups, and then transmitting one or more vector groups to the multilayer neural network.

The vector processing for converting the characteristic value into the vector value is an embedded technology, called Embedding, and is a technology for converting a discrete variable into a continuous vector in the field of neural networks.

The multilayer neural network processing unit is used for cross-combining word vectors in the vector group into data sets in the multilayer neural network, and constructing a shared data layer by using the data sets.

The multi-target model module is used for dividing data in the shared data layer constructed by the multi-layer neural network processing unit into a plurality of different data groups through a strategy.

The multi-target model module combines a plurality of different depolarization strategies to construct a universal depolarization model.

Dividing the data group is equivalent to dividing a plurality of target tasks based on different task purposes, different depolarization strategies are adopted for different data groups, and corresponding depolarization processing is equivalent to using the corresponding depolarization strategies for different task targets. The selection of the depolarization strategy based on each group of data is divided according to different task purposes, for example, a target task for eliminating the position deviation is established currently, and then the recalled data can be recalled according to the characteristic value of the position parameter of the input data when the input data is referred by using a tendency score method.

In order to ensure that the effect of data group division (namely the division of target tasks) is better, after the processing of the pre-order multi-layer neural network processing unit, the input data is subjected to multi-layer cross processing, the data dimensions of each piece of data in the shared data layer tend to be consistent, and at the moment, the data only need to be classified according to the target requirements; the specific data scale division can also be combined with the characteristic value input unit to divide the scale of the input data: the more data is divided by the multi-objective model, which type of data is input more.

And the depolarization model module is used for carrying out depolarization processing on the plurality of data groups divided by the multi-target model module by adopting a corresponding depolarization strategy, and integrating and outputting data after the depolarization processing of all the data groups.

The deskew model module integrates one or more processing units of different deskew strategies, for example, in this embodiment, the deskew strategy includes: a trend score method, a data filling method, a dual robust estimator, joint learning, etc.

Trend scoring refers to a class of statistical methods that use non-experimental or observed data for intervention effect analysis. The data filling method includes a mean value filling method and the like. A dual robust estimator refers to the use of an adaptive filter to select targets from clutter (other signals than the desired signal). Joint learning refers to the client training the model together under the coordination of the central server, while keeping decentralization and dispersion of the training data.

As shown in FIG. 3, the present invention also discloses a computer system for applying the above method, the computer system comprising a processor and a memory, the memory configured to store a series of computer-executable instructions and computer-accessible data associated with the series of computer-executable instructions.

When executed by a processor, the series of computer-executable instructions cause the processor to perform the method as described in the embodiment illustrated in FIG. 1 above.

Additionally, a non-transitory computer readable storage medium having stored thereon a series of computer executable instructions which, when executed by a computing device, cause the computing device to perform the method as described in the embodiment illustrated in fig. 1 above is also disclosed.

While, for purposes of simplicity of explanation, the methodologies are shown and described as a series of acts, it is to be understood and appreciated that the methodologies are not limited by the order of acts, as some acts may, in accordance with one or more embodiments, occur in different orders and/or concurrently with other acts from that shown and described herein or not shown and described herein, as would be understood by one skilled in the art.

Those of skill would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The various illustrative logical blocks, modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk (disk) and disc (disc), as used herein, includes Compact Disc (CD), laser disc, optical disc, Digital Versatile Disc (DVD), floppy disk and blu-ray disc where disks (disks) usually reproduce data magnetically, while discs (discs) reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.