阅读说明：本技术 一种基于自修复建模的反洗钱系统及其方法 (Self-repairing modeling based money laundering prevention system and method thereof ) 是由 俞书浩 于 2021-06-28 设计创作，主要内容包括：本发明涉及一种基于自修复建模的反洗钱系统及其方法,该系统包括依次连接的洗钱识别装置和洗钱告警装置,洗钱识别装置用于对输入的交易数据进行洗钱识别,并输出对应的识别结果给洗钱告警装置,洗钱识别装置包括依次连接的特征加工模块、建模模块、自修复模块和模型优化模块,特征加工模块用于从输入的交易数据中提取出对应的高维全量特征；建模模块用于根据高维全量特征,构建初版洗钱识别模型；自修复模块用于进行自修复闭环投产操作,持续得到反馈数据；模型优化模块用于根据反馈数据,对当前洗钱识别模型进行更新优化。与现有技术相比,本发明针对快速变化的洗钱场景,解决了投产后模型衰减的问题,同时不需要反复人工建模,节省人力资源。(The invention relates to an anti-money laundering system based on self-repairing modeling and a method thereof, wherein the system comprises a money laundering identification device and a money laundering alarm device which are sequentially connected, the money laundering identification device is used for carrying out money laundering identification on input transaction data and outputting a corresponding identification result to the money laundering alarm device, the money laundering identification device comprises a feature processing module, a modeling module, a self-repairing module and a model optimization module which are sequentially connected, and the feature processing module is used for extracting a corresponding high-dimensional full-scale feature from the input transaction data; the modeling module is used for constructing an initial money laundering identification model according to the high-dimensional full-scale features; the self-repairing module is used for carrying out self-repairing closed-loop production operation to continuously obtain feedback data; and the model optimization module is used for updating and optimizing the current money laundering identification model according to the feedback data. Compared with the prior art, the method solves the problem of model attenuation after production aiming at the money laundering scene with quick change, and simultaneously does not need repeated artificial modeling, thereby saving human resources.)

1. The anti-money laundering system based on self-repairing modeling is characterized by comprising a money laundering identification device (1) and a money laundering alarm device (2) which are sequentially connected, wherein the money laundering identification device (1) is used for carrying out money laundering identification on input transaction data and outputting a corresponding identification result to the money laundering alarm device (2), the money laundering identification device (1) comprises a feature processing module (11), a modeling module (12), a self-repairing module (13) and a model optimizing module (14) which are sequentially connected, and the feature processing module (11) is used for extracting corresponding high-dimensional full-quantity features from the input transaction data;

the modeling module (12) is used for constructing an initial money laundering identification model according to the high-dimensional full-scale features;

the self-repairing module (13) is used for performing self-repairing closed-loop production operation so as to continuously obtain feedback data;

the model optimization module (14) is used for updating and optimizing the current money laundering identification model according to the feedback data.

2. The self-healing modeling-based anti-money laundering system according to claim 1, wherein the feature processing module (11) comprises a data cleaning unit (110) and a data processing unit (111) connected in sequence, the data cleaning unit (110) is used for performing data cleaning on the input transaction data to screen out abnormal data;

and the data processing unit (111) generates corresponding high-dimensional full-scale features according to the cleaned transaction data.

3. The self-repairing modeling-based anti-money laundering system according to claim 1, wherein the self-repairing module (13) comprises an automatic modeling unit (130), a cloud commissioning unit (131) and a feedback data cleaning and processing unit (132) which are connected in a closed loop, and the feedback data cleaning and processing unit (132) is configured to acquire feedback data from the cloud commissioning unit (131) and perform cleaning and processing on the acquired feedback data to generate corresponding high-dimensional full-scale features;

the automatic modeling unit (130) automatically establishes a new model according to the high-dimensional full-scale characteristics output by the feedback data cleaning and processing unit (132);

and the cloud commissioning unit (131) is used for performing cloud deployment commissioning on the new model.

4. A self-healing modeling based anti-money laundering method applying the anti-money laundering system according to claim 1, comprising the steps of:

s1, setting a modeling target, inputting historical sample data into the feature processing module, and obtaining an initial money laundering identification model through the modeling module;

s2, inputting the transaction data to be identified into a money laundering identification device containing an initial money laundering identification model to output a corresponding money laundering identification result, and continuously optimizing the initial money laundering identification model through a self-repairing module and a model optimization module to continuously update the money laundering identification model in the money laundering identification device;

and S3, outputting corresponding warning information to the user by the money laundering warning device according to the money laundering identification result.

5. The self-repair modeling based anti-money laundering method according to claim 4, wherein the step S1 specifically comprises the steps of:

s11, setting a modeling target and acquiring historical sample data;

s12, performing data cleaning on the historical sample data;

s13, constructing a corresponding wide table according to the historical sample data after data cleaning, wherein the wide table contains full-scale features;

s14, carrying out data processing on the full-scale features in the wide table to obtain corresponding high-dimensional full-scale features;

and S15, establishing an original money laundering identification model based on the high-dimensional full-scale features obtained in the step S14.

6. The self-repair modeling based anti-money laundering method according to claim 5, wherein the modeling targets are specifically: money laundering transaction 1 and non-money laundering transaction 0.

7. The self-repairing modeling-based anti-money laundering method according to claim 5, wherein the step S15 is specifically to adopt an xgboost or lightgbm algorithm with addition of L0, L1 and L2 regularization to establish an original money laundering identification model, thereby solving the problems of overfitting and multiple collinearity generated when high-dimensional full-scale feature modeling is used.

8. The self-repair modeling based anti-money laundering method according to claim 4, wherein the process of updating the money laundering identification model in step S2 specifically comprises:

s21, inputting the transaction data to be identified into the current money laundering identification model, and outputting a corresponding money laundering identification result while obtaining corresponding feedback data;

s22, sequentially carrying out data cleaning and data processing on the feedback data to obtain corresponding high-dimensional full-scale features;

s23, according to the high-dimensional full-scale features obtained in the step S22, a new model is obtained through automatic modeling construction;

s24, carrying out cloud deployment and production on the new model, and then returning to the step S21;

and S25, updating and optimizing the current money laundering identification model according to the accumulated feedback data and a preset optimization period.

9. The self-repair modeling based anti-money laundering method according to claim 8, wherein the data processing is to perform sliding window statistics including 10 days, 1 month, 3 months and 6 months, and calculate the maximum value, the minimum value, the average value, the median value and the summation value.

10. The self-repair modeling based anti-money laundering method according to claim 8, wherein the step S23 is to construct a new model by using a bayesian stochastic automatic modeling algorithm.

Technical Field

The invention relates to the technical field of anti-money laundering, in particular to an anti-money laundering system based on self-repairing modeling and a method thereof.

Background

Along with the increase of the transaction amount of the banking business, the occurrence risk of the anti-money laundering case also increases gradually, the number of suspicious transactions to be checked screened in the anti-money laundering rule system per month is in an increasing trend, and the workload of the anti-money laundering work is increased. Reducing the number of suspicious cases only by manual rule optimization, it is difficult to establish a long-term mechanism of rule optimization, so using AI machine learning techniques to generate anti-money laundering models is becoming a research hotspot in the prior art.

However, since the money laundering means is changed frequently, the strategy is changed according to money laundering rules of banks, so that the behavior characteristics of money laundering are always changed, but the common anti-money laundering model is not well suitable for the change situation, so that the money laundering transaction cannot be accurately identified after the anti-money laundering model which is built in a period of several months is over one or two months, and the anti-money laundering transaction can only be repeatedly modeled manually, so that a large amount of human resources are consumed.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide an anti-money laundering system based on self-repairing modeling and a method thereof, so as to carry out model self-repairing through high-dimensional full-scale characteristics aiming at constantly changing money laundering behavior characteristics, thereby solving the problem of attenuation of an anti-money laundering model.

The purpose of the invention can be realized by the following technical scheme: an anti-money laundering system based on self-repairing modeling comprises a money laundering identification device and a money laundering alarm device which are sequentially connected, wherein the money laundering identification device is used for carrying out money laundering identification on input transaction data and outputting a corresponding identification result to the money laundering alarm device;

the modeling module is used for constructing an initial edition money laundering identification model according to the high-dimensional full-scale features;

the self-repairing module is used for performing self-repairing closed-loop production operation so as to continuously obtain feedback data;

and the model optimization module is used for updating and optimizing the current money laundering identification model according to the feedback data.

Furthermore, the characteristic processing module comprises a data cleaning unit and a data processing unit which are sequentially connected, wherein the data cleaning unit is used for cleaning the input transaction data to screen out abnormal data;

and the data processing unit generates corresponding high-dimensional full-scale features according to the cleaned transaction data.

Further, the self-repairing module comprises an automatic modeling unit, a cloud production unit and a feedback data cleaning and processing unit which are connected in a closed-loop manner, wherein the feedback data cleaning and processing unit is used for acquiring feedback data from the cloud production unit and cleaning and processing the acquired feedback data to generate corresponding high-dimensional full-scale features;

the automatic modeling unit is used for cleaning the high-dimensional full-quantity characteristics output by the processing unit according to the feedback data and automatically establishing a new model;

and the cloud commissioning unit is used for carrying out cloud deployment commissioning on the new model.

An anti-money laundering method based on self-repair modeling comprises the following steps:

s1, setting a modeling target, inputting historical sample data into the feature processing module, and obtaining an initial money laundering identification model through the modeling module;

s2, inputting the transaction data to be identified into a money laundering identification device containing an initial money laundering identification model to output a corresponding money laundering identification result, and continuously optimizing the initial money laundering identification model through a self-repairing module and a model optimization module to continuously update the money laundering identification model in the money laundering identification device;

and S3, outputting corresponding warning information to the user by the money laundering warning device according to the money laundering identification result.

Further, the step S1 specifically includes the following steps:

s11, setting a modeling target and acquiring historical sample data;

s12, performing data cleaning on the historical sample data;

s13, constructing a corresponding wide table according to the historical sample data after data cleaning, wherein the wide table contains full-scale features;

s14, carrying out data processing on the full-scale features in the wide table to obtain corresponding high-dimensional full-scale features;

and S15, establishing an original money laundering identification model based on the high-dimensional full-scale features obtained in the step S14.

Further, the modeling target is specifically: money laundering transaction 1 and non-money laundering transaction 0.

Further, the step S15 specifically adopts xgboost or lightgbm algorithm with regularization added to L0, L1, and L2 to build an original money laundering identification model, thereby solving the problems of overfitting and multiple collinearity generated when high-dimensional full-scale feature modeling is used.

Further, the process of updating the money laundering identification model in step S2 specifically includes:

s21, inputting the transaction data to be identified into the current money laundering identification model, and outputting a corresponding money laundering identification result while obtaining corresponding feedback data;

s22, sequentially carrying out data cleaning and data processing on the feedback data to obtain corresponding high-dimensional full-scale features;

s23, according to the high-dimensional full-scale features obtained in the step S22, a new model is obtained through automatic modeling construction;

s24, carrying out cloud deployment and production on the new model, and then returning to the step S21;

and S25, updating and optimizing the current money laundering identification model according to the accumulated feedback data and a preset optimization period.

Further, the data processing specifically includes performing sliding window statistics including 10 days, 1 month, 3 months, and 6 months, and calculating to obtain a maximum value, a minimum value, an average value, a median value, and a summation value.

Further, in the step S23, a new model is specifically constructed by using a bayesian stochastic automatic modeling algorithm.

Compared with the prior art, the invention has the following advantages:

the invention aims at money laundering scenes with fast change of money laundering behavior characteristics, and solves the problem of attenuation of money laundering identification models after production by arranging a characteristic processing module, a modeling module, a self-repairing module and a model optimizing module which are sequentially connected in a money laundering identification device, obtaining an original money laundering identification model by utilizing the characteristic processing module and the modeling module, and continuously repairing and updating the current money laundering identification model by utilizing the self-repairing module and the model optimizing module.

Secondly, the money laundering identification model is constructed based on the high-dimensional full-scale features, the high-dimensional full-scale features are used as the basis of self-repairing, input of each re-modeling is the same, and therefore subsequent machine learning modeling and commissioning can be performed fully automatically, after feedback data obtained after commissioning is processed, the data format of the feedback data is consistent with the original high-dimensional full-scale features, manual participation is not needed, and therefore the self-repairing quick updating performance of the money laundering identification model is guaranteed.

Drawings

FIG. 1 is a schematic diagram of the system of the present invention;

FIG. 2 is a schematic flow diagram of the process of the present invention;

FIG. 3 is a schematic diagram of the operation of the money laundering identification device;

the notation in the figure is: 1. the system comprises a money laundering identification device, 2, a money laundering alarm device, 11, a characteristic processing module, 12, a modeling module, 13, a self-repairing module, 14, a model optimization module, 110, a data cleaning unit, 111, a data processing unit, 130, an automatic modeling unit, 131, a cloud production unit, 132 and a feedback data cleaning processing unit.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.

Examples

As shown in fig. 1, an anti-money laundering system based on self-repairing modeling includes a money laundering identification device 1 and a money laundering alarm device 2 which are connected in sequence, the money laundering identification device 1 is used for performing money laundering identification on input transaction data and outputting a corresponding identification result to the money laundering alarm device 2, the money laundering identification device 1 includes a feature processing module 11, a modeling module 12, a self-repairing module 13 and a model optimization module 14 which are connected in sequence, the feature processing module 11 is used for extracting a corresponding high-dimensional full-scale feature from the input transaction data, the feature processing module 11 includes a data cleaning unit 110 and a data processing unit 111 which are connected in sequence, the data cleaning unit 110 is used for performing data cleaning on the input transaction data to screen out abnormal data;

the data processing unit 111 generates corresponding high-dimensional full-scale features according to the cleaned transaction data;

the modeling module 12 is used for constructing an initial money laundering identification model according to the high-dimensional full-scale features;

the self-repairing module 13 is used for performing self-repairing closed-loop production operation to continuously obtain feedback data, the self-repairing module 13 comprises an automatic modeling unit 130, a cloud production unit 131 and a feedback data cleaning and processing unit 132 which are connected in a closed-loop manner, and the feedback data cleaning and processing unit 132 is used for acquiring feedback data from the cloud production unit 131 and cleaning and processing the acquired feedback data to generate corresponding high-dimensional full-scale features;

the automatic modeling unit 130 automatically establishes a new model according to the high-dimensional full-scale features output by the feedback data cleaning and processing unit 132;

the cloud commissioning unit 131 is configured to perform cloud deployment commissioning on the new model;

the model optimization module 14 is configured to perform update optimization on the current money laundering identification model according to the feedback data.

The above system is applied to practice to realize an anti-money laundering method based on self-repairing modeling, as shown in fig. 2, and comprises the following steps:

s1, setting a modeling target, inputting historical sample data into the feature processing module 11, and obtaining an initial money laundering identification model through the modeling module 12, specifically:

firstly, setting a modeling target: money laundering transaction 1 and non-money laundering transaction 0, and obtaining historical sample data;

then, data cleaning is carried out on the historical sample data;

then, according to historical sample data after data cleaning, constructing a corresponding wide table, wherein the wide table comprises full-scale features;

finally, data processing is carried out on the full-scale features in the wide table to obtain corresponding high-dimensional full-scale features, and an initial money laundering identification model is established based on the obtained high-dimensional full-scale features;

s2, inputting the transaction data to be identified into the money laundering identification device 1 including the first-edition money laundering identification model, so as to output a corresponding money laundering identification result, and meanwhile, continuously optimizing the first-edition money laundering identification model through the self-repairing module 13 and the model optimizing module 14, and continuously updating the money laundering identification model in the money laundering identification device 1, specifically:

s21, inputting the transaction data to be identified into the current money laundering identification model, and outputting a corresponding money laundering identification result while obtaining corresponding feedback data;

s22, sequentially carrying out data cleaning and data processing on the feedback data to obtain corresponding high-dimensional full-scale features, wherein the data processing specifically comprises carrying out sliding window statistics of 10 days, 1 month, 3 months and 6 months, and calculating to obtain a maximum value, a minimum value, an average value, a median value and a summation value;

s23, according to the high-dimensional full-scale features obtained in the step S22, a new model is obtained through automatic modeling construction;

s24, carrying out cloud deployment and production on the new model, and then returning to the step S21;

s25, updating and optimizing the current money laundering identification model according to the accumulated feedback data and a preset optimization period;

and S3, outputting corresponding warning information to the user by the money laundering warning device 2 according to the money laundering identification result.

According to the technical scheme, the traditional anti-money laundering model can be automatically and iteratively optimized according to feedback data, but when characteristics are generated in a transaction flow, due to the problem of computing power, if modeling is performed only by simply counting characteristics such as transaction stroke number, transaction amount and the like, effective characteristics are screened, and then a finally optimized model is obtained through machine learning to put into production, due to the change of behavior characteristics, a plurality of old characteristics can be eliminated and need new characteristics to support, so that under the mode of optimizing the model, the model accuracy rate can still be rapidly reduced.

Therefore, according to the technical scheme, a self-repairing anti-money laundering modeling mode is realized, the modeling process is improved, the problem of model attenuation after production is solved, the problem of rapid response to changes is solved, repeated manual modeling is not needed, a large amount of human resources are saved, a large amount of high-dimensional features are automatically generated on the basis of common anti-money laundering modeling, screening of the features is not needed, and self-repairing can be realized after the model effect is attenuated.

Based on the traditional anti-money laundering modeling, high-dimensional full-scale features are generated, and a restoration ring is established by using an initial version model with a common effect. The repair loop comprises four parts, namely 'high-dimensional full-scale feature', 'automatic modeling algorithm', 'model automatic commissioning' and 'feedback data', so that a closed loop is formed.

As shown in fig. 3:

1. feature processing module

Including demand analysis, data preparation, data cleansing and exploration, and simple feature engineering to form broad tables. Firstly, according to a target, behavior data of a mobile phone client is collected, then data processing is carried out, abnormal data are screened out, and a high-dimensional full-scale feature is generated, in the embodiment, a modeling target is specified: anti-money laundering transactions and non-anti-money laundering transactions (1, 0); then preparing data, cleaning the data, finding abnormal data and improving the data quality; generating a high-dimensional full-scale feature: and 300 main features, wherein each main feature is subjected to sliding window statistics for 10 days, 1 month, 3 months and 6 months in sequence, and the maximum value, the minimum value, the average value, the median value and the summation value are calculated.

2. Modeling module

After the wide table is completed, the initial model can be modeled for the first time to obtain an initial model (or a baseline model), and the initial model can try several algorithms and select the best one of the algorithms; then, aiming at the selected specific algorithm, model parameters are optimized, feature screening is carried out, and the effect of the model is adjusted to a satisfactory degree, in the embodiment, the self-repairing anti-money laundering modeling technology is considered to use high-dimensional full-scale features, so that overfitting is easily caused, and therefore an algorithm capable of resisting overfitting and multiple collinearity is needed, and therefore L0, L1 and L2 regularization are added on the basis of the open-source xgboost and lightgbm algorithms, so that the problems of overfitting and multiple collinearity are effectively solved.

3. Self-repairing module

The self-repairing is based on 'high-dimensional full-quantity characteristics', and because the input of each modeling is different due to characteristic screening in the common method, manual intervention is needed for each re-modeling; and by using the high-dimensional full-quantity feature, the input of each re-modeling is the same, and the subsequent processes, namely the automatic machine learning algorithm and the automatic production can be automatically executed. After the 'feedback data' obtained after production is processed, the data format of the 'feedback data' is consistent with the 'high-dimensional full-scale feature' of the original edition, so that manual participation is not needed, and the method mainly comprises the following steps: (1) the feedback data are automatically cleaned, the high-dimensional full-scale features generated by the step (2) are automatically deployed in a cloud deployment mode by adopting a Bayesian random automatic modeling algorithm (3).

4. Model optimization module

After the self-repairing anti-money laundering modeling process is completed, models of initial versions do not pursue high accuracy, but the models can accumulate new 'feedback data' during operation to gradually improve, and the accuracy is improved; when the accuracy is stable, the model can quickly cope with the change, and the rapid attenuation is avoided. The whole modeling process embedded with self-repairing only needs one-time development, the problem of model attenuation is not worried after production, the difficult problem of rapid response to change is solved, and meanwhile, repeated artificial modeling is not needed, so that a large amount of human resources are saved.

In the technical scheme, the models of the initial versions do not pursue high accuracy, but the models can accumulate new 'feedback data' during operation to gradually improve, so that the accuracy is improved; when the accuracy is stable, the model can quickly cope with the change, and the rapid attenuation is avoided. The whole self-repairing anti-money laundering modeling process only needs one-time development, the problem of model attenuation is not needed to be worried about after production, the problem of rapid response to changes is solved, and meanwhile, repeated manual modeling is not needed, so that a large amount of human resources are saved. The traditional model has a good effect when just put into production, but quickly attenuates after 1 month, then manpower is put into the model after attenuation, and the new model is the same fate after 3 months. The self-repairing model may not have the effect as the common model at first, but then the model gradually improves along with the data accumulation, the accuracy rate gradually rises, and then the high accuracy rate can be kept all the time by the quick repairing which depends on the self-repairing model. The model updating period of the common method needs 3 months of manpower, while the model updating period using self-repairing can be very short, can be 2 weeks or 1 week, and does not need manual intervention.

According to the technical scheme, a self-repairing modeling mode is adopted for a fast-changing money laundering scene, automatic closed-loop repair is realized on the basis of high-dimensional full-quantity characteristics, production is carried out through automatic cloud deployment, repeated manual modeling is not needed, and a large amount of human resources are saved.