Risk control method, model training method, equipment and storage medium

Abstract

The embodiment of the invention provides a risk control method, a model training method, equipment and a storage medium. Acquiring user data of a target user after a request of applying for a target service by the target user is received; judging whether the user data is abnormal or not; if the user data is abnormal, correcting the abnormity in the user data to obtain reconstructed data without the abnormity; and performing risk control on the behavior of applying for the target service by the target user based on the reconstruction data. Before risk control is carried out on a behavior of applying for a target service by a user based on user data, a link of identifying abnormal data and eliminating anomalies in the data can be added, and risk control can be carried out based on more real and accurate reconstructed data in a subsequent risk control link by restoring the abnormal data into an original feature, so that the risk control efficiency is improved. And the risk control of the target business is more accurate.

Description

Technical Field

[0001] Embodiments of this specification relate to the field of computer technology, and in particular, to a risk control method, a model training method, a device, and a storage medium. Background Art

[0002] With the development of the Internet, users can achieve various financial services through the Internet. For example, users can apply for various services such as loans, claims, and financing through the Internet. Considering that in the above service scenarios, there may be some risky user application behaviors (such as fraudulent application behaviors, users without repayment ability applying for loans, etc.), these risky application behaviors will cause losses to financial institutions providing funds. Therefore, in such service scenarios, it is necessary to control the risks of users' behaviors of applying for these services to reduce the losses of financial institutions.

[0003] In related technologies, usually, a pre-trained model can be used to control the risks of users' behaviors of applying for the above services. For example, it can be determined whether a user has fraudulent behavior, or whether a user is a high-risk user, etc., to determine whether to provide funds to the user and the amount of funds provided. Currently, usually, user data is directly input into a pre-trained model, and risk decisions are made through the model. Although this method can effectively identify most of the risky application behaviors and conduct control, for some situations where user data is abnormal due to objective reasons (such as large fluctuations in data, user data obtained from upstream business systems not being updated in time, etc.), it will also be determined as a risky application behavior, and the applications of these users will be rejected. It can be seen that in related technologies, it is not yet possible to accurately determine the risks existing in users' behaviors of applying for the above services for accurate risk control, and there are some misjudgment problems, which will seriously affect the user experience. Summary of the Invention

[0004] To overcome the problems existing in related technologies, embodiments of this specification provide a risk control method, a model training method, a device, and a storage medium.

[0005] According to a first aspect of the embodiments of this specification, a risk control method is provided, and the method includes:

[0006] After receiving a request from a target user to apply for a target service, obtain the user data of the target user;

[0007] Determine whether the user data is abnormal;

[0008] If the user data is abnormal, correct the abnormality in the user data based on the association relationship between different types of user data of the target user and / or the association relationship between the user data of the target user and the user data of other users, to obtain reconstructed data with the abnormality eliminated;

[0009] Perform risk control on the behavior of the target user applying for the target service based on the reconstructed data.

[0010] According to the second aspect of the embodiments of this specification, a model training method is provided, and the method includes:

[0011] Obtain an original sample, where the original sample includes original user data in one or more dimensions;

[0012] Add perturbation information to the original sample through the generator of the generative adversarial model to generate an adversarial sample;

[0013] Input the adversarial sample and the original sample into the discriminator of the generative adversarial model respectively, and distinguish the adversarial sample and the original sample through the discriminator;

[0014] Adjust the model parameters of the generative adversarial model based on the difference between the determination result of the discriminator and the true result, to train the generative adversarial model.

[0015] According to the third aspect of the embodiments of this specification, a model training method is provided, and the method includes:

[0016] Obtain an adversarial sample, where the adversarial sample is generated by adding interference information to an original sample, and the original sample includes original user data in one or more dimensions;

[0017] Extract the sample features of the adversarial sample by using the encoder in the autoencoder model;

[0018] Use the decoder in the autoencoder model to reconstruct a reconstructed sample with the perturbation eliminated based on the sample features;

[0019] Adjust the model parameters of the autoencoder model based on the difference between the reconstructed sample and the original sample, to train the autoencoder model.

[0020] According to the fourth aspect of the embodiments of this specification, an electronic device is provided, and the electronic device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, and when the computer program is executed, it implements the methods mentioned in the first aspect, the second aspect, and / or the third aspect above.

[0021] According to a fifth aspect of the embodiments of this specification, a computer storage medium is provided. A computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the methods mentioned in the above first aspect, second aspect, and / or third aspect are implemented.

[0022] According to a sixth aspect of the embodiments of this specification, a computer program product is provided. The computer program product includes a computer program, and when the computer program is executed by a processor, the methods mentioned in the above first aspect, second aspect, and / or third aspect are implemented.

[0023] The beneficial effects of the embodiments of this specification are as follows: When performing risk control on a user's behavior of applying for a target service based on user data, it is possible to first determine whether the user data is abnormal. After determining that the user data is abnormal, the abnormality in the user data can be corrected to obtain reconstructed data with the abnormality eliminated, and then risk control can be performed on the user's behavior of applying for the target service based on the reconstructed data with the abnormality eliminated. The solution of the embodiments of this specification can add a link of identifying abnormal data and eliminating the abnormality in the data before performing risk control on the user's behavior of applying for the target service based on user data. By restoring the abnormal data to its original appearance, in the subsequent risk control link, risk control can be performed based on more real and accurate reconstructed data, making the risk control of the target service more accurate and reducing the misjudgment phenomenon.

[0024] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the embodiments of this specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings herein are incorporated into the specification and constitute a part of the embodiments of this specification, showing embodiments that conform to the embodiments of this specification, and are used together with the specification to explain the principles of the embodiments of this specification.

[0026] Figure 1 It is a flowchart of a risk control method shown in an exemplary embodiment of this specification;

[0027] Figure 2 It is a schematic diagram of the training process of a generative adversarial model shown in an exemplary embodiment of this specification;

[0028] Figure 3 It is a schematic diagram of the training process of an autoencoder model shown in an exemplary embodiment of this specification;

[0029] Figure 4 It is a schematic diagram of the processing flow of an Internet peer-to-peer lending business in the related art;

[0030] Figure 5A schematic diagram of the processing flow of the Internet peer-to-peer lending business shown in an exemplary embodiment of this specification;

[0031] FIG. 6(a) is a schematic diagram for determining whether a user is a risky user in the related art;

[0032] FIG. 6(b) is a schematic diagram for determining whether a user is a risky user shown in an exemplary embodiment of this specification;

[0033] Figure 7 A logic block diagram of an electronic device shown in an exemplary embodiment of this specification. Detailed implementation manners

[0034] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the embodiments of this specification. On the contrary, they are merely examples of devices and methods consistent with some aspects of the embodiments of this specification as detailed in the appended claims.

[0035] The terms used in the embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of this specification. The singular forms "a", "the" and "said" used in the embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0036] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of this specification to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the embodiments of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0037] With the development of the Internet, users can realize various financial services through the Internet. For example, users can apply for loans, claim compensations, financing and other various services through the Internet. Considering that in the above business scenarios, there may be some risky user application behaviors (such as fraudulent application behaviors, users without repayment ability applying for loans, etc.), these risky application behaviors will cause losses to the financial institutions providing funds. Therefore, in such business scenarios, it is necessary to perform risk control on the application behaviors of these services to reduce the losses of financial institutions.

[0038] In related technologies, the behavior of a user applying for the above-mentioned services can usually be risk-controlled through a pre-trained model. For example, it can be determined whether a user has fraudulent behavior, whether a user is a high-risk user, etc., so as to determine whether to provide funds to the user and the amount of funds provided. Currently, usually the user data is directly input into the pre-trained model, and the risk decision is made through the model. Although this method can effectively identify most of the application behaviors with risks and conduct control, for some situations where the user data is abnormal due to objective reasons, it will also be determined as an application with risks and the applications of these users will be rejected.

[0039] For example, taking the fraudulent behavior of a user as an example, illegal users may forge data or tamper with data to achieve the purpose of insurance fraud or loan fraud, thus causing losses to the financial institutions providing the above-mentioned services. In order to accurately identify the fraudulent behavior of users, usually a fraud detection model is used to detect the fraud of users applying for the above-mentioned services. For example, it can be determined whether a user has fraudulent behavior based on the user data of the user. If there is no fraudulent behavior, the subsequent process will be carried out. Although this method can effectively identify most of the user's fraudulent behaviors and reject these service applications, there are also some problems.

[0040] For example, generally, there may be two situations where the user data is abnormal: (1) The user data is abnormal due to the user maliciously tampering with or forging data; (2) The user data is abnormal due to some objective reasons such as large fluctuations in the data or the user data obtained from the upstream business system not being updated in time. Currently, when the fraud detection model conducts fraud detection, usually for those situations where the user data is abnormal, it is directly determined as a fraudulent behavior, thus rejecting the user application of this user. The current processing method is likely to determine the situation where the user data is abnormal due to objective reasons as a fraudulent behavior and reject the applications of such users, affecting the user experience.

[0041] Of course, the situation is similar for other risk control scenarios. In related technologies, the accuracy and reliability of the obtained user data are not considered, and the risk control is directly based on the user data, so it is impossible to accurately determine the risks existing in the behavior of a user applying for the above-mentioned services, and thus it is impossible to accurately control the risks. There are still some misjudgment problems, seriously affecting the user experience.

[0042] Based on this, the embodiments of this specification provide a risk control method. When performing risk control on the behavior of a user applying for a target service based on user data, it is possible to first determine whether the user data is abnormal. After determining that the user data is abnormal, the abnormality in the user data can be corrected to obtain reconstructed data with the abnormality eliminated. For example, for an abnormality caused by a user's tampering behavior, the user data can be restored to the data before being tampered with. For an abnormality caused by the user data obtained not being updated in a timely manner, the user data can be restored to the updated data. Then, risk control can be performed on the behavior of the user applying for the target service based on the reconstructed data after eliminating the abnormality.

[0043] That is, before performing risk control on the behavior of a user applying for a target service based on user data, the solution of the embodiments of this specification can add a link to identify abnormal data and eliminate the abnormality in the data. By restoring the abnormal data to its original appearance, in the subsequent risk control link, risk control can be performed based on more real and accurate reconstructed user data, making the risk control of the target service more precise and reducing the phenomenon of misjudgment.

[0044] The risk control method of the embodiments of this specification can be executed by various electronic devices. For example, the electronic device can be a terminal, a server, or a server cluster used to implement the target service, etc.

[0045] As Figure 1 shown, the risk control method provided by the embodiments of this specification can include the following steps:

[0046] S102. After receiving a request from a target user to apply for a target service, obtain the user data of the target user;

[0047] In step S102, when receiving a request from a target user to apply for a target service, the user data of the target user can be obtained. Among them, the target service can be various Internet financial services. For example, it can be Internet loans, Internet insurance, Internet financing, etc. The user data of the target user can be various data related to the target service. For example, it can be the attribute data of the user, such as age, gender, salary status, credit data, IP address, MAC address, etc.; the historical behavior data of the user, such as the user's historical loan behavior, repayment behavior, consumption behavior, etc.; the relationship network graph data between the user and other users, etc.

[0048] Among them, the user data can be the user data provided by the user together when submitting the application request, or the user data obtained by the business system processing the target service from other business systems through various legal and compliant channels. For example, the credit data of the user can be obtained from the credit system. The embodiments of this specification do not make any restrictions.

[0049] S104. Determine whether there is an abnormality in the user data;

[0050] In the related art, after obtaining the user data of the target user, the risk control of the user's application behavior is directly based on the user data. For example, determining whether the user has fraudulent behavior, determining whether the user is a risky user, determining the loan amount of the user, etc. This method does not consider the accuracy and reliability of the user data itself, resulting in inaccurate processing results for subsequent various tasks.

[0051] Therefore, in step S104, after obtaining the user data of the target user, the accuracy and reliability of the user data itself can be determined first to determine whether there is an abnormality in the user data. Among them, the user data abnormality means that the user data does not match the actual situation, or the user data deviates from the normal change trend. For example, taking the user age as an example, assuming that based on other data of the user, the user is determined to be a minor user (i.e., less than 18 years old), and the currently obtained user age is much greater than 18 years old, it means that there is an abnormality in the user age data item.

[0052] In some embodiments, the user data abnormality may be an abnormality caused by the user's data forgery behavior, or an abnormality caused by various objective reasons. For example, it may be that the user deliberately tampered with and forged data in order to defraud funds, or it may be that there are indeed large fluctuations in the data itself, or the obtained data is data that has not been updated in time, resulting in the data deviating from the normal change trend. In this step, the user data abnormality caused by the above two reasons can be identified.

[0053] Among them, determining whether there is an abnormality in the user data can be determined by various methods. For example, it can be determined by a specially trained model, or by means of correlation analysis and comparison of the data.

[0054] S106. If there is an abnormality in the user data, correct the abnormality in the user data based on the correlation relationship between different types of user data of the target user, and / or the correlation relationship between the user data of the target user and the user data of other users;

[0055] In step S106, if it is determined that there is an abnormality in the user data, the abnormality in the user data can be corrected to obtain reconstructed data with the abnormality eliminated. Among them, correcting the abnormality in the user data is to restore the original appearance, original change rule or trend of the data. For example, for the abnormality caused by the user's tampering behavior, the user data can be restored to the data before being tampered with or close to the data before being tampered with. For the abnormality caused by the fact that the obtained user data is data that has not been updated in time, the user data can be restored to the updated data.

[0056] Among them, when correcting the anomalies in user data, the anomalies in user data can be corrected based on the association relationships between different types of user data of the target user. For example, the internal relationships or rules between different types of user data of the same user can be learned, and the normal user data can be predicted based on such internal rules or relationships. For example, the normal salary of a user can be predicted based on the provident fund data paid by the user to correct the abnormal salary.

[0057] Of course, when correcting the anomalies in user data, the anomalies in user data can also be corrected based on the association relationships between the user data of different users. For example, the internal rules and relationships of the user data between different users can be learned, and the normal user data can be predicted based on such internal rules or relationships. For example, by analyzing the consumption data of a large number of minors, the characteristics of the consumption behavior of minors can be summarized. If the consumption behavior of a certain user conforms to such characteristics, but the age in the user data submitted by him is the age of an adult, therefore, his age can be corrected and determined to be a minor.

[0058] For example, taking Internet loans as an example, it is usually stipulated that loan services cannot be provided to minors. Some minor users may forge data and tamper with their ages in order to successfully obtain a loan. For example, if the actual age of a user is 16 years old, but his age is forged to 28 years old. In this case, it can be determined whether the age of the user is normal based on other information of the user. If it is determined based on other information that the user is actually a minor, it is considered that there is an anomaly in the user's age. At this time, the age of the user can be re-determined based on other information of the user, so that the reconstructed age is close to the true age of the user, and then it can be determined whether to provide a loan to the user based on the reconstructed age.

[0059] Another example is taking the salary data of a user as an example. It may be determined that the currently obtained salary data of the user is inaccurate based on the user's provident fund and social security payment situations. In related technologies, it may be directly considered that the user has committed fraud, and thus the user's application is rejected. Then, in some scenarios, this salary data may be salary data that has not been updated in time, rather than being caused by tampering. Therefore, the salary data of the user can be re-determined based on the user's provident fund and social security payment situations to obtain the salary that conforms to the true situation of the user, and then it can be determined whether to provide a loan to the user based on the reconstructed salary data.

[0060] Among them, correcting the anomalies in user data to eliminate the anomalies in user data can be achieved through a pre-trained model, or by analyzing the change trends of different types of user data of the same user, the user data between different users, or the association relationships between user data, and predicting the data after eliminating the anomalies based on such change trends and association relationships.

[0061] S108. Perform risk control on the behavior of the target user applying for the target service based on the reconstructed data.

[0062] In step S108, after correcting the anomalies in the user data to obtain the reconstructed data with anomalies eliminated, risk control can be performed on the application behavior of the target user for the target service based on the reconstructed data. Among them, performing risk control on the application behavior of the target user for the target service can be determining whether there is fraud in the application behavior (for example, tampering with or forging data) to determine whether to reject the application of the target user, determining whether the target user is a high-risk user to make a risk decision on the application behavior, determining the amount of funds corresponding to the application behavior (for example, loan amount, claim amount, etc.), or one or more of them.

[0063] In some embodiments, if it is determined that there are no anomalies in the user data of the target user, risk control can be directly performed on the application behavior of the target user for the target service based on the user data.

[0064] In some embodiments, it can be determined whether there are anomalies in the user data through the discriminator of a pre-trained generative adversarial model. Among them, the generative adversarial model can be trained based on multiple original samples, and the original samples include original user data in one or more dimensions.

[0065] A generative adversarial network (GAN) is a deep learning framework that trains a model through a game. A generative adversarial model consists of two parts: a generator and a discriminator, which compete and cooperate with each other during training. During the training process, the generator can generate new samples similar to the real samples, and the discriminator can distinguish between the real samples and the new samples generated by the generator and output a prediction result indicating the probability that the input sample is a real sample. Among them, the training goal of the generator is to generate new samples as close as possible to the real samples so that the generated new samples can be misidentified as real samples by the discriminator. The training goal of the discriminator is to accurately distinguish between real samples and the new samples generated by the generator, and the two constantly compete with each other, thereby obtaining a generator and a discriminator with relatively high accuracy.

[0066] In some embodiments, such as Figure 2As shown, when training the generative adversarial model, original samples can be obtained. The original samples include original user data in one or more dimensions, where the original user data includes various types of data related to the target business, such as the user's attribute data, the user's historical behavior data, and the user's relationship network graph data with other users. Then, the original samples can be input into the generator of the generative adversarial model, and the generator of the generative adversarial model adds perturbation information to the original samples to generate adversarial samples. Then, the adversarial samples and the original samples are respectively input into the discriminator of the generative adversarial model, and the discriminator distinguishes between the adversarial samples and the original samples. For example, it determines whether the input sample is an original sample or an adversarial sample generated by the generator. Then, based on the difference between the determination result of the discriminator and the true result, the model parameters of the generative adversarial model can be adjusted to train the generative adversarial model.

[0067] When adding perturbation information to the original samples to generate adversarial samples, the original user data in one or more dimensions of the original samples can be perturbed. Among them, the perturbation methods can include multiple types. For example, the data in one or more dimensions can be modified in a specific manner to make the perturbed data change towards the expected target, or the data in one or more dimensions can be randomly modified to make the perturbed data change randomly, or a unit vector with the same dimension as the original sample can be directly added to the original sample to obtain the perturbed data.

[0068] In some embodiments, in order to enable the discriminator in the finally trained generative adversarial model to accurately identify the anomalies caused by the user's data forgery behavior and the anomalies caused by objective reasons among the two situations. When the generator adds perturbation information to the original samples to generate adversarial samples, the way of adding perturbation information can simulate the above two abnormal phenomena, so that the obtained adversarial samples can cover the anomalies in the above two situations. For example, perturbation rules can be preset, and the generator can modify the original user data in one or more dimensions according to the preset perturbation rules to generate adversarial samples. Among them, the perturbation rules include the first type of perturbation rules for simulating the user's data forgery behavior and the second type of perturbation rules for simulating the data anomaly phenomena caused by objective reasons.

[0069] For example, it is possible to analyze and count the data forgery behaviors of users, summarize the characteristics of the forged data of users, and then formulate perturbation rules based on these characteristics to generate abnormal data simulating the data forgery behaviors of users. Of course, for data anomalies caused by objective reasons such as data fluctuations and untimely data updates, it is also possible to summarize the characteristics of such abnormal data, and then formulate perturbation rules based on these characteristics to simulate the situation of abnormal data caused by objective reasons. When training the discriminator of the generative adversarial model, by generating adversarial samples covering the above two types of anomalies to train the discriminator, the finally trained discriminator can accurately identify the anomalies caused by these two types of reasons.

[0070] In some embodiments, the anomalies in user data can be corrected through a pre-trained autoencoder model to obtain reconstructed data with anomalies eliminated. Among them, the autoencoder model can be trained by using adversarial samples as inputs and the original samples corresponding to the adversarial samples as labels. Among them, the adversarial samples are samples generated by adding interference information to the original samples.

[0071] Generally, an autoencoder model includes two parts: an encoder and a decoder. The encoder is used to extract the features of the input data, that is, compress the input data into a low-dimensional vector, usually called a code, and transfer it to the decoder. The decoder receives the code generated by the encoder and performs reverse operations based on this code to reconstruct new data.

[0072] In some embodiments, as Figure 3 shown, when training the autoencoder model, multiple sets of training data can be obtained. Among them, each set of training includes an adversarial sample and the original sample corresponding to the adversarial sample. Among them, the original sample includes original user data in one or more dimensions, and the adversarial sample is generated by adding perturbation information to the original sample. The original sample can be used as the label of the adversarial sample to train the autoencoder model. During the training process, the adversarial sample can be first input into the encoder of the autoencoder model, and the encoder extracts the features of the adversarial sample to obtain the sample features of the adversarial sample. Then the sample features can be input into the decoder of the autoencoder model, and the decoder can perform reverse operations based on the sample features to reconstruct a reconstructed sample with perturbations eliminated. Then, based on the difference between the reconstructed sample and the original sample, the model parameters of the autoencoder model can be adjusted to train the autoencoder model. Among them, the training objective is to enable the autoencoder model to reconstruct a reconstructed sample as close as possible to the original sample of the input adversarial sample, so that the finally trained autoencoder model can reconstruct a more accurate original sample based on the adversarial sample and restore the original appearance of the original data as accurately as possible.

[0073] In some embodiments, considering that during the process of training an autoencoder model, the sample features extracted by the encoder from adversarial samples may be high-dimensional features. If the dimension of the sample features is larger, that is, the adversarial samples are characterized in more detail, then the decoder may over-learn the features of the adversarial samples, that is, there is a problem of overfitting, resulting in the reconstructed samples finally reconstructed by the decoder not being accurate enough. To reduce the above-mentioned overfitting problem, after using the encoder to extract features from the adversarial samples to obtain sample features, the sample features can be first subjected to dimensionality reduction processing, that is, mapping the sample features from a high-dimensional space to a low-dimensional space, mapping the sample features into features with a lower dimension, and obtaining the sample features after dimensionality reduction. For example, assuming that the extracted sample features are a vector in a high-dimensional space, that is, the vector can be mapped into a vector in a low-dimensional space to perform dimensionality reduction processing on the sample features. Then the sample features after dimensionality reduction can be input into the decoder, and the decoder constructs a reconstructed sample based on the sample features after dimensionality reduction processing. By performing dimensionality reduction processing on the sample features and then inputting them into the decoder, it is possible to prevent the decoder from over-learning the characteristics of the sample features and prevent the problem of overfitting.

[0074] In some embodiments, the adversarial samples for training the autoencoder model can be generated based on the generator of a pre-trained generative adversarial model. Among them, when training the generative adversarial model, original samples can be obtained, where the original samples include original user data in one or more dimensions. Then the original samples can be input into the generator of the generative adversarial model, and the generator of the generative adversarial model adds perturbation information to the original samples to generate adversarial samples. Then the adversarial samples and the original samples are respectively input into the discriminator of the generative adversarial model, and the discriminator distinguishes between the adversarial samples and the original samples. For example, it determines whether the input sample is an original sample or an adversarial sample generated by the generator, and then the model parameters of the generative adversarial model can be adjusted based on the difference between the determination result of the discriminator and the true result to train the generative adversarial model. Among them, the specific training method of the generative adversarial model can refer to the description in the above embodiments and will not be elaborated here.

[0075] In some embodiments, when performing risk control on the behavior of a target user applying for the target service based on the reconstructed data, the reconstructed data can be input into a pre-trained anti-fraud model. The anti-fraud model is used to determine whether the behavior of the target user applying for the target service is a fraudulent behavior. If it is, the application request of the target user is directly rejected. If not, subsequent processing steps are carried out. Among them, the anti-fraud model can be trained based on the historical application behavior data of the user for the target service. For example, the historical application behavior data of the user for the target service can be obtained, and these data are used to construct training samples. For example, for each group of user data, it can be determined whether the label corresponding to this group of user data is "fraudulent behavior" or "non-fraudulent behavior". Then, the preset model can be trained based on these user data with labels to obtain the above anti-fraud model.

[0076] In some embodiments, the target service can be an Internet peer-to-peer lending service. The Internet lending service generally includes three links: anti-fraud detection, risk detection, and credit limit approval. The three links can be implemented through three pre-trained models (anti-fraud model, risk determination model, credit limit approval model), and the types of user data used in the three links are not completely the same. As Figure 4 shown, in the related art, when a loan application from a user is received, the first type of user data used to determine whether the user has fraudulent behavior can be obtained from the user data and input into the anti-fraud model. The anti-fraud model determines whether there is fraudulent behavior. If there is, the loan application of the user is rejected. Otherwise, the second type of user data used to determine the risk level of the user is obtained and input into the risk determination model to determine whether the user is a risky user. If so, the user request is rejected. Otherwise, the third type of user data used to determine the loan amount of the user is obtained, and then input into the credit limit approval model to determine the loan amount of the user, and then the user is credited based on the loan amount.

[0077] In the above three links, the user data used may all be abnormal (this abnormality can be caused by the user forging data or by objective reasons). If the obtained user data is directly used for the processing of the above three links, there will be a problem that all situations of abnormal user data are determined as fraudulent behaviors, affecting the user experience. Moreover, if the anti-fraud model in the first link does not accurately identify fraudulent behavior, the subsequent risk determination and credit limit approval links will continue to use the abnormal data for risk determination and loan amount determination, resulting in inaccurate processing results for all three links.

[0078] In addition, in the related art, the three processes of anti-fraud detection, risk detection, and credit limit approval are carried out independently. However, the user data used in the three processes often has an internal connection, and the current processing method ignores this connection.

[0079] In the embodiments of this specification, for the scenario where a target user applies for an Internet P2P lending service, the user data of the target user can be first determined for anomalies. If there are no anomalies, subsequent anti-fraud detection, risk detection, and credit limit approval can be directly carried out. If the user data has anomalies, the user data can be first processed for anomaly correction to obtain reconstructed data with anomalies eliminated. Based on the reconstructed data corresponding to the first type of user data, it can be determined whether the target user has fraud behavior. If there is fraud behavior, the application request of the target user can be directly rejected. If there is no fraud behavior, based on the reconstructed data corresponding to the second type of user data, it can be determined whether the target user is a risky user. If it is determined that the target user is a risky user, the application request of the target user is rejected. If it is determined that the target user is not a risky user, the loan amount of the target user is determined based on the reconstructed data corresponding to the third type of user data, and the account of the target user is credited with the loan amount. As Figure 5 shown, the above three processes can also be implemented through three pre-trained models (anti-fraud model, risk determination model, credit limit approval model).

[0080] Among them, the first type of user data can be various data used to determine whether a user has fraud behavior. For example, the user's attribute data, the user's social network data, the user's time-point data, and so on. Similarly, the second type of user data can be various data used to determine whether a user is a risky user. For example, the user's historical loan behavior, consumption behavior, repayment record, credit investigation data, and so on. The third type of user data can be various data used to determine the loan amount of a user. For example, the user's provident fund, social security data, the user's consumption level, the user's qualification data, etc.

[0081] Through the solution provided by the embodiments of this specification, on the one hand, in the anti-fraud process, anti-fraud determination can be carried out based on the first type of user data with anomalies eliminated, so that it is possible to avoid determining all cases of user data anomalies as fraud behavior, reduce misjudgment phenomena, and improve the user experience. On the other hand, since anomalies are eliminated for all user data, that is, deviation correction, more authentic and reliable data is obtained. When subsequent risk determination is carried out based on the reconstructed second type of user data and the loan amount is determined based on the reconstructed third type of user data, the results obtained are also more accurate.

[0082] In addition, for the scenario of determining whether there are abnormalities in user data during the training of the generative adversarial model and reconstructing user data using the autoencoder model, since the above three types of user data are used in the training processes of the two models in the early stage, that is, during the training processes of the two models, the associations between these three types of user data are learned. As a result, the internal relationships between the above three types of user data are also considered in the finally reconstructed data. By processing the following three tasks based on the reconstructed data, more accurate processing results can also be obtained.

[0083] In the related art, for Internet credit services, in the risk detection link, as shown in Fig. 6(a), usually, the behavior data of a user at a certain time point (for example, consumption behavior data), the basic portrait of the user, and credit data and other data are input into a risk determination model (hereinafter, the above data are collectively referred to as traditional data), and the probability P that the user is a risky user is output through this risk determination model. The data used in this way is not comprehensive enough, and thus the finally obtained risk detection result is also inaccurate.

[0084] In order to improve the accuracy of the risk determination result, in some embodiments, the second type of user data for determining whether a target user is a risky user may include, in addition to the above traditional data, the temporal behavior data of the target user (for example, the consumption behavior data and loan behavior data of the user within a certain time period, etc.), the text data related to the user (for example, user reviews, user addresses, etc.), and the social network data of the user. When determining whether the target user is a risky user based on the reconstructed data corresponding to the second type of user data, a temporal model may be used to determine the first probability that the target user is a risky user based on the reconstructed data of the temporal behavior data, a language model may be used to determine the second probability that the target user is a risky user based on the reconstructed data of the text data, and a graph neural network model may be used to determine the third probability that the target user is a risky user based on the reconstructed data of the social network data. Then, based on the first probability, the second probability, the third probability, and their respective weights, a target probability is obtained, and based on the target probability, it is determined whether the target user is a risky user.

[0085] For example, as shown in Figure 6(b), traditional data such as a user's behavioral data (e.g., consumption behavioral data), basic portrait, and credit investigation data at a certain time point can be input into a risk determination model. Through this risk determination model, the probability P1 that the user is a risky user is output. Then, text data related to the user can be output to a language model, and through this language model, the probability P2 that the user is a risky user is output. Then, the user's sequential behavioral data can be input into a sequential model, such as the NHFM model (Neural Factorization Machine). Through the sequential model, the probability P3 that the user is a risky user is output. Then, the user's social network data is input into a graph neural network model, and through this model, the probability P4 that the user is a risky user is output. Furthermore, a comprehensive probability P can be obtained based on the probabilities P1, P2, P3, and P4 output by each model. Among them, for the above four probabilities, respective corresponding weights can be set, and the comprehensive probability P is obtained based on their respective weights. Based on this comprehensive probability P, it is determined whether the user is a risky user. For example, the comprehensive probability P can be obtained by using the multi-dimensional model Boosting method.

[0086] In the embodiments of this specification, by also considering the user's sequential behavioral data, text-related data, and social network data, and then using models good at processing these data to predict the probability that the user is a risky user based on these data respectively, and comprehensively determining whether the user is a risky user based on these probabilities, the accuracy of the determination result can be improved.

[0087] In addition, this application also provides a model training method, and the method includes:

[0088] Obtain an original sample, where the original sample includes original user data in one or more dimensions;

[0089] Add perturbation information to the original sample through the generator of the generative adversarial model to generate an adversarial sample;

[0090] Input the adversarial sample and the original sample into the discriminator of the generative adversarial model respectively, and the discriminator distinguishes the adversarial sample and the original sample;

[0091] Based on the difference between the determination result of the discriminator and the true result, adjust the model parameters of the generative adversarial model to train the generative adversarial model.

[0092] Among them, the specific details of the above model training method can refer to the description in the above embodiments and will not be elaborated here.

[0093] Furthermore, this application also provides a model training method, and the method includes:

[0094] Obtain adversarial samples, where the adversarial samples are generated by adding interference information to the original samples, and the original samples include original user data in one or more dimensions;

[0095] Use the encoder in the autoencoder model to extract the sample features of the adversarial samples;

[0096] Use the decoder in the autoencoder model to reconstruct a reconstructed sample without perturbations based on the sample features;

[0097] Adjust the model parameters of the autoencoder model based on the difference between the reconstructed sample and the original sample to train the autoencoder model.

[0098] Wherein, the specific details of the above model training method can be referred to the description in the above embodiments and will not be elaborated here.

[0099] In addition, an embodiment of this specification also provides a computer program product, which includes a computer program that implements the method mentioned in any of the above embodiments when executed by a processor. Corresponding to the embodiment of the risk control method provided in this specification, this specification also provides a risk control device, and the device includes:

[0100] An acquisition module, configured to acquire the user data of the target user after receiving a request from the target user to apply for a target service;

[0101] A determination module, configured to determine whether the user data is abnormal;

[0102] A reconstruction module, configured to correct the abnormality in the user data based on the association relationship between different types of user data of the target user and / or the association relationship between the user data of the target user and the user data of other users if the user data is abnormal; a risk control module, configured to perform risk control on the behavior of the target user applying for the target service based on the reconstructed data.

[0103] The implementation processes of the functions and roles of each unit in the above device are specifically detailed in the implementation processes of the corresponding steps in the above method and will not be elaborated here.

[0104] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the descriptions of the method embodiments. The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the objectives of the embodiments of this specification. A person of ordinary skill in the art can understand and implement it without creative work.

[0105] In terms of the hardware level, as Figure 7 shown, it is a hardware structure diagram of the device where the risk control device of the embodiments of this specification is located. In addition to Figure 7 the processor 72 and the memory 74 shown, this device usually may also include other hardware, such as a forwarding chip responsible for processing packets, etc.; in terms of the hardware structure, this device may also be a distributed device, and may include multiple interface cards to expand packet processing at the hardware level. The memory 74 stores computer instructions, and when the processor 72 executes the computer instructions, it implements the methods mentioned in any of the above embodiments.

[0106] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. And the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions, and corresponding operation entrances are provided for users to choose to authorize or refuse.

[0107] Since the part of the embodiments of this specification that contributes to the prior art or all or part of this technical solution can be embodied in the form of a software product, this computer software product is stored in a storage medium and includes several instructions to enable a terminal device to execute all or part of the steps of the methods of the embodiments of this specification. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0108] The above are only the preferred embodiments of the embodiments of this specification and are not intended to limit the embodiments of this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of this specification shall be included within the scope of protection of the embodiments of this specification.

Claims

1. A risk control method, the method comprising: After receiving a request from a target user for a target service, obtaining user data of the target user; Determining whether the user data is abnormal; If the user data is abnormal, correcting the abnormality in the user data based on the association relationship between different types of user data of the target user, and / or the association relationship between the user data of the target user and the user data of other users, to obtain reconstructed data with the abnormality eliminated; Performing risk control on the behavior of the target user applying for the target service based on the reconstructed data.

2. The method according to claim 1, wherein the determination of whether the user data is abnormal is implemented by a discriminator of a pre-trained generative adversarial model, and the generative adversarial model is trained based on the following method: Obtaining an original sample, the original sample including original user data in one or more dimensions; Adding perturbation information to the original sample through a generator of the generative adversarial model to generate an adversarial sample; Respectively inputting the adversarial sample and the original sample into the discriminator of the generative adversarial model, and distinguishing the adversarial sample and the original sample through the discriminator; Adjusting the model parameters of the generative adversarial model based on the difference between the determination result of the discriminator and the true result to train the generative adversarial model.

3. The method according to claim 2, wherein the abnormality is caused by the user's data forgery behavior, and / or the abnormality is caused by an objective reason; The step of adding perturbation information to the original sample through a generator of a preset generative adversarial model to generate an adversarial sample includes: Modifying the original user data in one or more dimensions in the original sample according to a preset perturbation rule to generate an adversarial sample, wherein the perturbation rule includes a first type of perturbation rule for simulating the user's data forgery behavior and a second type of perturbation rule for simulating data abnormality phenomena caused by objective reasons.

4. The method according to claim 1, wherein the correction of the abnormality in the user data based on the association relationship between different types of user data of the target user, and / or the association relationship between the user data of the target user and the user data of other users, to obtain reconstructed data with the abnormality eliminated is implemented by a pre-trained autoencoder model, and the autoencoder model is trained based on the following method: Obtaining an adversarial sample, the adversarial sample being generated by adding interference information to an original sample, the original sample including original user data in one or more dimensions; Extracting sample features of the adversarial sample by using an encoder in the autoencoder model; Reconstructing a reconstructed sample with the perturbation eliminated based on the sample features by using a decoder in the autoencoder model; Adjusting the model parameters of the autoencoder model based on the difference between the reconstructed sample and the original sample to train the autoencoder model.

5. The method according to claim 4, before reconstructing a reconstructed sample with the perturbation eliminated based on the sample features by using a decoder in the autoencoder model, the method further includes: Map the sample features from a high-dimensional space to a low-dimensional space to obtain sample features with reduced dimensions; The using the decoder in the autoencoder model to reconstruct a reconstructed sample without perturbations based on the sample features includes: Reconstructing a reconstructed sample without perturbations based on the sample features with reduced dimensions.

6. The method according to claim 4, wherein the adversarial sample is generated based on a generator of a pre-trained generative adversarial model, and the generative adversarial model is trained based on the following method: Obtain an original sample, where the original sample includes original user data in one or more dimensions; Add perturbation information to the original sample through the generator of the generative adversarial model to generate an adversarial sample; Input the adversarial sample and the original sample into the discriminator of the generative adversarial model respectively, and distinguish the adversarial sample and the original sample through the discriminator; Adjust the model parameters of the generative adversarial model based on the difference between the determination result of the discriminator and the true result to train the generative adversarial model.

7. The method according to claim 1, wherein the method further comprises: If the user data is normal, directly perform risk control on the target service applied for by the target user based on the user data; and / or The performing risk control on the behavior of the target user applying for the target service based on the reconstructed data includes: inputting the reconstructed data into a pre-trained anti-fraud model, and determining whether the target user has fraudulent behavior through the anti-fraud model; wherein, the anti-fraud model is trained based on historical application behavior data of the user for the target service; If there is fraudulent behavior, reject the request of the target user to apply for the target service.

8. The method according to claim 1, wherein the target service is an Internet peer-to-peer lending service, and the user data includes: The first type of user data for determining whether the target user has fraudulent behavior, the second type of user data for determining whether the target user is a risk user, and the third type of user data for determining the loan amount of the target user, The performing risk control on the behavior of the target user applying for the target service based on the reconstructed data includes: Determining whether the target user has fraudulent behavior based on the reconstructed data corresponding to the first type of user data; If there is no fraudulent behavior, determining whether the target user is a risk user based on the reconstructed data corresponding to the second type of user data; If not, determining the loan amount of the target user based on the reconstructed data corresponding to the third type of user data, and transferring the funds to the account of the target user based on the loan amount.

9. The method according to claim 8, wherein the second type of user data at least includes the time-series behavior data of the target user, the text data related to the target user, and the social network data of the target user; The determining whether the target user is a risk user based on the reconstructed data corresponding to the second type of user data includes: Using a time-series model to determine the first probability that the target user is a risk user based on the reconstructed data corresponding to the time-series behavior data; Using a language model to determine the second probability that the target user is a risk user based on the reconstructed data corresponding to the text data; Determine a third probability that the target user is a risky user based on the reconstructed data corresponding to the social network data by using a graph neural network model; Obtain a target probability based on the first probability, the second probability, the third probability, and their respective corresponding weights, and determine whether the target user is a risky user based on the target probability.

10. A model training method, the method comprising: Obtain an original sample, the original sample including original user data in one or more dimensions; Add perturbation information to the original sample through a generator of a generative adversarial model to generate an adversarial sample; Input the adversarial sample and the original sample into a discriminator of the generative adversarial model respectively, and distinguish the adversarial sample and the original sample through the discriminator; Adjust model parameters of the generative adversarial model based on the difference between the determination result of the discriminator and the true result to train the generative adversarial model.

11. A model training method, the method comprising: Obtain an adversarial sample, the adversarial sample being generated by adding interference information to an original sample, the original sample including original user data in one or more dimensions; Extract sample features of the adversarial sample by using an encoder in an autoencoder model; Reconstruct a reconstructed sample without perturbation based on the sample features by using a decoder in the autoencoder model; Adjust model parameters of the autoencoder model based on the difference between the reconstructed sample and the original sample to train the autoencoder model.

12. A computer program product, the computer program product including a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1-11 above is implemented.

13. An electronic device, the electronic device including a processor, a memory, and a computer program stored on the memory and executable by the processor, and when the processor executes the computer program, the method according to any one of claims 1-11 above can be implemented.

14. A computer storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method according to any one of claims 1-11 above is implemented.

Images