Method for constructing multi-scene credit risk control model and risk control processing method and device

Abstract

The invention relates to a method for constructing a multi-scene credit risk control model and a risk control processing method and device, and the method comprises the steps: obtaining training data, which comprises user credit related historical data and credit states in a plurality of target scenes; performing feature extraction and feature preprocessing on the historical data related to the user credit to obtain training sample features; the training sample features are input into a to-be-trained multi-task learning model for training, default prediction probabilities in multiple target scenes are output, and the training labels are credit states in the corresponding target scenes; the trained multi-task learning model is used as a multi-scene credit risk control model; wherein the multi-task learning model is used for performing feature cross processing on input to obtain output features fused with global context information; and carrying out default prediction processing in the plurality of target scenes based on the output features to obtain corresponding default prediction probabilities in the target scenes. The model can give consideration to multi-scene prediction and guarantee the prediction accuracy.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular, to a method for constructing a multi-scenario credit risk control model, a method for risk control processing, and an apparatus therefor. Background Art

[0002] In the financial field or related fields, such as scenarios of corporate financing, personal borrowing, credit rating for the establishment or operation of financial institutions, etc., there is a need to evaluate the credit status of enterprises, individuals, financial institutions, etc.

[0003] In the process of implementing the concept of the present disclosure, the inventors found that there are at least the following technical problems in the related art: In the related art, most of them use some machine learning models such as scorecards, some binary classification models, etc. to perform credit risk prediction in a single scenario; while in the risk control scenario, there is often a need to perform credit risk prediction for multiple scenarios; for this need, it is often difficult to make the risk control model take into account multi-scenario prediction and ensure prediction accuracy. Summary of the Invention

[0004] To solve the above technical problems or at least partially solve the above technical problems, embodiments of the present disclosure provide a method for constructing a multi-scenario credit risk control model, a method for risk control processing, and an apparatus therefor.

[0005] In a first aspect, an embodiment of the present disclosure provides a method for constructing a multi-scenario credit risk control model. The method includes: obtaining training data, where the training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the credit statuses are used to indicate whether a user defaults in each target scenario; performing feature extraction and feature preprocessing on the historical data related to user credit to obtain training sample features; inputting the training sample features into a multi-task learning model to be trained for training, where the output of the multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training labels are the credit statuses in the corresponding target scenarios; using the trained multi-task learning model as the multi-scenario credit risk control model; wherein, the multi-task learning model is used to: perform feature cross-processing on the input to obtain output features fused with global context information; and perform default prediction processing for multiple target scenarios based on the output features to obtain the corresponding default prediction probabilities in each target scenario.

[0006] According to an embodiment of the present disclosure, the above multi-task learning model includes: a feature embedding layer, a feature cross layer, a multi-task learning layer, a plurality of tower layers corresponding to a plurality of target scenarios, and a plurality of output layers corresponding to the plurality of target scenarios. The above feature embedding layer is used to perform feature embedding processing on the above training sample features to obtain embedding features of a preset dimension. The above feature cross layer is used to perform feature cross processing on the above embedding features to obtain output features fused with global context information. The above multi-task learning layer is used to perform default prediction learning for the above output features under a plurality of target scenarios to obtain learning results under each target scenario. Each tower layer has a fully connected layer structure, and the input of each tower layer is the learning result under the corresponding target scenario. Each output layer has a fully connected layer structure, the input of each output layer is the output result of the corresponding tower layer, and each output layer outputs the default prediction probability under the corresponding target scenario.

[0007] According to an embodiment of the present disclosure, the above training sample features include: m features corresponding to a user, and after corresponding feature embedding processing, m embedding features are obtained, where m≥2 and is an integer; the above feature cross layer includes k feature cross modules arranged in parallel, and each feature cross module includes: a feature guiding network, a normalization network, and a feed-forward network, where k≥2 and is an integer; the above feature guiding network includes: an integration layer and a projection layer, the above integration layer is used to perform dimension expansion processing on the input m embedding features and output a first feature; the above projection layer is used to perform dimension reduction processing on the above first feature and output a second feature; the dimensions of dimension expansion by the respective integration layers in the k feature cross modules are different; the above normalization network is used to perform normalization processing on the input m embedding features and output normalized features; perform an element-wise product operation (also called a Hadamard product operation) on the above second feature and the above normalized features to obtain a cross output result; the above feed-forward network is used to process the above cross output result and output a fused feature fused with global context information corresponding to each dimension expansion dimension, and the above fused feature is used as the output of each feature cross module; splice the fused features output by the k feature cross modules to obtain a k-dimensional output feature.

[0008] According to an embodiment of the present disclosure, the above-mentioned training sample features include: a total of m features in k feature domains corresponding to a user, and after corresponding feature embedding processing, m embedded features are obtained, where m≥2 and is an integer, k≥2 and is an integer; the total number of features in the i-th feature domain among the above k feature domains is mi, and the value of i ranges from 1 to k. The above-mentioned feature crossing layer is a two-layer structure. The first layer includes k feature crossing modules arranged in parallel, and the second layer includes r feature crossing modules arranged in parallel, where r≥2 and is an integer; each feature crossing module in the first layer includes: a first feature guiding network, a first normalization network, and a first feed-forward network; each feature crossing module in the second layer includes: a second feature guiding network, a second normalization network, and a second feed-forward network; alternatively, the first layer includes k feature crossing modules arranged in parallel, and each feature crossing module includes: a feature guiding network, a normalization network, and a feed-forward network, and the second layer is a fully connected layer structure. In the i-th feature crossing module of the first layer, the above-mentioned first feature guiding network includes: a first integration layer and a first projection layer; the above-mentioned first integration layer is used to perform dimension expansion processing on the input mi embedded features and output a third feature; the above-mentioned first projection layer is used to perform dimension reduction processing on the above-mentioned third feature and output a fourth feature; the above-mentioned first normalization network is used to perform normalization processing on the input mi embedded features and output a first normalized feature; perform an element-wise product operation on the above-mentioned third feature and the above-mentioned first normalized feature to obtain a first cross-output result; the above-mentioned first feed-forward network is used to process the above-mentioned first cross-output result and output a first fusion feature corresponding to the mi embedded features, and the above-mentioned first fusion feature is used as the output of the i-th feature crossing module; splice the first fusion features output by the k feature crossing modules in the first layer to obtain a first layer output feature, and the above-mentioned first layer output feature is used to be input into the feature crossing modules in the second layer. In each feature crossing module of the second layer, the above-mentioned second feature guiding network includes a second integration layer and a second projection layer; the above-mentioned second integration layer is used to perform dimension expansion processing on the input k first layer output features to obtain a fifth feature; the above-mentioned second projection layer is used to perform dimension reduction processing on the above-mentioned fifth feature and output a sixth feature; the dimensions of the second integration layers in the r feature crossing modules in the second layer for dimension expansion are different; the above-mentioned second normalization network is used to perform normalization processing on the input k first layer output features and output a second normalized feature; perform an element-wise product operation on the above-mentioned sixth feature and the above-mentioned second normalized feature to obtain a second cross-output result; the above-mentioned second feed-forward network is used to process the above-mentioned second cross-output result and output a second fusion feature; splice the second fusion features output by the r feature crossing modules in the second layer to obtain the output feature of the above-mentioned feature crossing layer.

[0009] According to an embodiment of the present disclosure, the above multi-task learning layer is one of the following structures: a multi-gate mixture-of-experts (MMoE) structure, a progressive layered extraction (PLE) structure.

[0010] According to an embodiment of the present disclosure, feature extraction and feature preprocessing are performed on the above historical data related to user credit to obtain training sample features, including: based on a feature extraction model, performing feature extraction on various types of data in the above historical data related to user credit to obtain multiple types of features; for each type of feature in the above multiple types of features, calculating at least one of the following indexes of the current feature: missing rate, information value (iv) value, sample stability index (psi) value, discrimination evaluation index (ks) value, positive and negative sample ratio evaluation index (woe) value; screening the above multiple types of features according to the sample screening index conditions and the indexes calculated from the above multiple types of features to obtain target features for subsequent training; performing at least one of the following preprocessing operations on the above target features: performing data deduplication processing, filling missing values with 0, performing one-hot encoding processing on discrete features, and performing Z-Score normalization processing on continuous features to obtain training sample features.

[0011] In a second aspect, an embodiment of the present disclosure provides a method for multi-scenario user credit granting risk control processing. The above method includes: obtaining credit-related data of a user to be evaluated; performing feature extraction and feature preprocessing on the above credit-related data to obtain user features; inputting the above user features into a pre-constructed multi-scenario credit granting risk control model, and outputting the default prediction probability of the above user to be evaluated in multiple target scenarios; wherein, the above multi-scenario credit granting risk control model is constructed by using the method for constructing a multi-scenario credit granting risk control model as described above.

[0012] In a third aspect, embodiments of the present disclosure provide an apparatus for constructing a multi-scenario credit risk control model. The apparatus includes: a first data acquisition module, a first feature extraction and preprocessing module, and a training module. The first data acquisition module is configured to acquire training data, where the training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the credit statuses are used to indicate whether a user defaults in each target scenario. The first feature extraction and preprocessing module is configured to perform feature extraction and feature preprocessing on the historical data related to user credit to obtain training sample features. The training module is configured to input the training sample features into a multi-task learning model to be trained for training, where the output of the multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training labels are the credit statuses in the corresponding target scenarios; the trained multi-task learning model is used as the multi-scenario credit risk control model. Among them, the multi-task learning model is configured to: perform feature cross-processing on the input to obtain output features fused with global context information; and perform default prediction processing in multiple target scenarios based on the output features to obtain the corresponding default prediction probabilities in each target scenario.

[0013] In a fourth aspect, embodiments of the present disclosure provide an apparatus for multi-scenario user credit risk control processing. The apparatus includes: a second data acquisition module, a second feature extraction and preprocessing module, and a processing module. The second data acquisition module is configured to acquire credit-related data of a user to be evaluated. The second feature extraction and preprocessing module is configured to perform feature extraction and feature preprocessing on the credit-related data to obtain user features. The processing module is configured to input the user features into a pre-constructed multi-scenario credit risk control model and output the default prediction probabilities of the user to be evaluated in multiple target scenarios; among them, the multi-scenario credit risk control model is constructed by using the method for constructing a multi-scenario credit risk control model as described above.

[0014] In a fifth aspect, embodiments of the present disclosure provide an electronic device. The electronic device includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus; the memory is used to store a computer program; the processor is configured to, when executing the program stored in the memory, implement the method for constructing a multi-scenario credit risk control model or the method for multi-scenario user credit risk control processing as described above.

[0015] In a sixth aspect, embodiments of the present disclosure provide a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the method for constructing a multi-scenario credit risk control model or the method for multi-scenario user credit risk control processing as described above.

[0016] At least some or all of the advantages of the technical solutions provided by the embodiments of the present disclosure are as follows:

[0017] By extracting and preprocessing features from historical data related to user credit, training sample features are obtained; the preprocessing process can improve the data quality of the training sample features, which helps to improve the training efficiency of the multi-task learning model and the stability and accuracy of the multi-scenario credit risk control model; by integrating multi-task learning and feature crossing, feature crossing processing is performed on the input to obtain output features fused with global context information, and default prediction processing is performed on the above output features in multiple target scenarios to obtain the corresponding default prediction probabilities in each target scenario. Feature crossing processing can not only obtain the interaction information between global features, but also assign higher weights to important features, making important features more prominent and reducing the influence of weak features and noise. Multi-task learning can cover multiple target scenarios at the same time, effectively avoiding the cumbersome process of modeling and parameter adjustment for multiple single scenarios, and realizing risk control prediction for multiple target scenarios in the same model and effectively improving the prediction accuracy in multiple target scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present disclosure and used together with the specification to explain the principles of the present disclosure.

[0019] To more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the accompanying drawings required for use in the description of the embodiments or related technologies. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 Schematically shows the system architecture of a method for building a multi-scenario credit risk control model and a method for multi-scenario user credit risk control processing applicable to the embodiments of the present disclosure;

[0021] Figure 2 Schematically shows the flowchart of a method for building a multi-scenario credit risk control model according to an embodiment of the present disclosure;

[0022] Figure 3A Schematically shows the structural diagram of a multi-task learning model according to an embodiment of the present disclosure;

[0023] Figure 3B Schematically shows the structural diagram of a multi-task learning model according to another embodiment of the present disclosure;

[0024] Figure 4 Schematically shows the flowchart of a method for multi-scenario user credit risk control processing according to an embodiment of the present disclosure;

[0025] Figure 5 Schematically shows a structural block diagram of an apparatus for constructing a multi-scenario credit risk control model according to an embodiment of the present disclosure;

[0026] Figure 6 Schematically shows a structural block diagram of an apparatus for multi-scenario user credit risk control processing according to an embodiment of the present disclosure;

[0027] Figure 7 Schematically shows a structural block diagram of an electronic device provided by an embodiment of the present disclosure. Detailed implementation manners

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.

[0029] Figure 1 Schematically shows a system architecture applicable to the methods for constructing a multi-scenario credit risk control model and multi-scenario user credit risk control processing according to embodiments of the present disclosure.

[0030] Refer to Figure 1 As shown, the system architecture 100 applicable to the methods for constructing a multi-scenario credit risk control model and multi-scenario user credit risk control processing according to embodiments of the present disclosure includes: terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is a medium providing a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0031] Users can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various client applications, such as financial applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples), may be installed on the terminal devices 101, 102, 103.

[0032] The terminal devices 101, 102, 103 may be various electronic devices with a display screen. For example, the electronic devices include but are not limited to: smart phones, tablet computers, laptop computers, desktop computers, smart watches, smart bracelets, in-vehicle smart devices, etc.

[0033] Server 105 may be a server that provides various services. For example, it can be a background management server (only for illustration) that provides service support for financial applications accessed by users using terminal devices 101, 102, and 103. The background management server can analyze and process the received financial data processing requests, and feedback the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0034] It should be noted that the methods for constructing a multi-scenario credit risk control model and the methods for multi-scenario user credit risk control provided in the embodiments of the present disclosure can generally be executed by server 105 or a terminal device with certain computing capabilities. Correspondingly, the devices for constructing a multi-scenario credit risk control model and the devices for multi-scenario user credit risk control provided in the embodiments of the present disclosure can generally be set in server 105 or the above-mentioned terminal device with certain computing capabilities. The methods for constructing a multi-scenario credit risk control model and the methods for multi-scenario user credit risk control provided in the embodiments of the present disclosure can also be executed by a server or a server cluster that is different from server 105 and can communicate with at least one of terminal devices 101, 102, 103 or server 105. Correspondingly, the devices for constructing a multi-scenario credit risk control model and the devices for multi-scenario user credit risk control provided in the embodiments of the present disclosure can also be set in a server or a server cluster that is different from server 105 and can communicate with at least one of terminal devices 101, 102, 103 or server 105.

[0035] It should be understood, Figure 1 the numbers of the terminal devices, the network, and the servers in

[0036] The first exemplary embodiment of the present disclosure provides a method for constructing a multi-scenario credit risk control model.

[0037] Figure 2 Schematically shows a flowchart of a method for constructing a multi-scenario credit risk control model according to an embodiment of the present disclosure.

[0038] Referring to Figure 2 as shown, the method for constructing a multi-scenario credit risk control model provided in the embodiments of the present disclosure includes the following steps: S210, S220, and S230.

[0039] In step S210, training data is obtained. The above training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the above credit statuses are used to indicate whether the user defaults in each target scenario.

[0040] In some embodiments, the multiple target scenarios may be, but are not limited to, the following scenarios: multiple default prediction scenarios corresponding to different prediction periods, default prediction scenarios corresponding to different credit products, default prediction scenarios corresponding to different credit-granting stages, etc.

[0041] The user credit-related historical data may be the user's basic information obtained after the user's authorization, such as, but not limited to: income information, industry information, relevant application usage data, device type information, consumption behavior data, payment data, etc.

[0042] The credit status corresponding to the same user under multiple target scenarios may be a set of status data associated with the target scenarios. For example, the credit status corresponding to Scenario A is: no default; the credit status corresponding to Scenario B is: default; the credit status corresponding to Scenario C is: no default.

[0043] In step S220, feature extraction and feature preprocessing are performed on the above user credit-related historical data to obtain training sample features.

[0044] According to an embodiment of the present disclosure, in the above step S220, performing feature extraction and feature preprocessing on the above user credit-related historical data to obtain training sample features includes: based on a feature extraction model, performing feature extraction on various types of data in the above user credit-related historical data to obtain multiple types of features; for each type of feature in the above multiple types of features, calculating at least one of the following indexes of the current feature: missing rate, information value iv value, sample stability index psi value, discrimination evaluation index ks value, positive and negative sample ratio evaluation index woe value; screening the above multiple types of features according to the sample screening index conditions and the indexes calculated from the above multiple types of features to obtain target features for subsequent training; performing at least one of the following preprocessing operations on the above target features: performing data deduplication processing, filling missing values with 0, performing one-hot encoding processing on discrete features, and performing Z-Score normalization processing on continuous features to obtain training sample features.

[0045] In some embodiments, it is also possible to remove field information such as user ID and time in the above user credit-related historical data that is irrelevant to training.

[0046] For example, the sample screening index conditions are at least one of the following conditions: removing samples with a missing rate greater than 0.9, removing samples with an iv value less than 0.01, and removing samples with a psi value greater than 0.1. By setting the sample screening index conditions and screening the multiple types of features, the feature dimension for training can be reduced, the training efficiency can be improved, and at the same time, it is also helpful to improve the stability and accuracy of the multi-scenario credit-granting risk control model obtained through training.

[0047] By performing Z-Score normalization on continuous features, the influence of the unit and scale differences between features is eliminated, making the process of finding the optimal solution smoother and easier to converge when training the multi-task learning model subsequently.

[0048] In step S230, the above-mentioned training sample features are input into the multi-task learning model to be trained. The output of the above-mentioned multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training label is the credit status in the corresponding target scenario; the trained multi-task learning model is used as the multi-scenario credit risk control model; wherein, the above-mentioned multi-task learning model is used to: perform feature cross-processing on the input to obtain output features fused with global context information; and perform default prediction processing in multiple target scenarios based on the above-mentioned output features to obtain the corresponding default prediction probabilities in each target scenario.

[0049] Figure 3A Schematically shows a schematic structural diagram of a multi-task learning model according to an embodiment of the present disclosure; Figure 3B Schematically shows a schematic structural diagram of a multi-task learning model according to another embodiment of the present disclosure.

[0050] According to an embodiment of the present disclosure, referring to Figure 3A and Figure 3B As shown, the above-mentioned multi-task learning model 300 includes: a feature embedding layer 310, a feature cross layer 320, a multi-task learning layer 330, a plurality of tower layers (TowerLayer) 340 corresponding to multiple target scenarios, and a plurality of output layers (Output Layer) 350 corresponding to multiple target scenarios. For example, the number of target scenarios is n, n≥2 and is an integer, and the number of corresponding tower layers and output layers is also n.

[0051] The above-mentioned feature embedding layer (Embedding Layer) 310 is used to perform feature embedding processing on the above-mentioned training sample features to obtain embedding features of a preset dimension.

[0052] The above-mentioned feature cross layer 320 is used to perform feature cross-processing on the above-mentioned embedding features to obtain output features fused with global context information.

[0053] The above-mentioned multi-task learning layer 330 is used to perform default prediction learning in multiple target scenarios on the above-mentioned output features to obtain the learning results in each target scenario.

[0054] Each tower layer is a fully connected layer structure, and the input of each tower layer is the learning result in the corresponding target scenario.

[0055] Each output layer is a fully connected layer structure, the input of each output layer is the output result of the corresponding tower layer, and each output layer outputs the default prediction probability in the corresponding target scenario.

[0056] According to an embodiment of the present disclosure, the above-mentioned training sample features include: m features corresponding to a user, and after corresponding feature embedding processing, m embedding features are obtained, where m≥2 and is an integer. Refer to Figure 3A As shown, the above-mentioned feature crossing layer includes k feature crossing modules arranged in parallel. In Figure 3A each, the respective feature crossing modules are denoted as mask_block1 to mask_block k. Each feature crossing module includes: an instance-guided mask, a normalization network (LN_EMB), and a feed-forward network. The feed-forward network includes a hidden layer network Hidden Layer and a normalization network LN_HID, where k≥2 and is an integer. The above-mentioned instance-guided mask includes: an aggregation layer and a projection layer. The above-mentioned aggregation layer is used to perform dimension expansion processing on the input m embedding features and output a first feature; the above-mentioned projection layer is used to perform dimension reduction processing on the above-mentioned first feature and output a second feature; the dimensions of dimension expansion by the respective aggregation layers in the k feature crossing modules mask_block1 to mask_blockk are different. The above-mentioned normalization network (LN_EMB) is used to perform normalization processing on the input m embedding features and output normalized features. Refer to Figure 3A As shown, an element-wise product operation (also known as a Hadamard product operation) is performed on the above-mentioned second feature and the above-mentioned normalized feature to obtain a cross-output result. The above-mentioned feed-forward network is used to process the above-mentioned cross-output result and output a fused feature with global context information corresponding to each dimension of dimension expansion. The above-mentioned fused feature is used as the output of each feature crossing module. The fused features output by the k feature crossing modules are concatenated to obtain a k-dimensional output feature.

[0057] In this embodiment, the multi-task learning model 300 includes a plurality of feature crossing modules arranged in parallel, and mainly constructs cross features based on the element-wise product method. During the process of constructing cross features, the embedding features corresponding to all inputs are interactively learned, which can fuse global context information, enhance effective features and weaken noise. The above-mentioned aggregation layer and projection layer are two fully connected layer structures. By expanding the dimension of the input and then reducing the dimension, the global context information can be fully utilized from the input samples, and the dimensions of dimension expansion by the respective aggregation layers in the k feature crossing modules mask_block 1 to mask_block k are different. The number of neurons in the aggregation layer is usually greater than the length of the input to better capture the feature information of the global context. The parameters of this layer are Wd1 The parameters of the projection layer are W d2 The input of the feature guidance network is the output V of the first-layer Embedding emb , and the output result is:

[0058] V mask = W d2 (W d1 V emb + β d1 ) + β d2 , (1)

[0059] The result of the feature crossing method based on element-wise product is:

[0060] V MaskedEMB = V mask * LN(V emb ), (2)

[0061] where LN represents normalizing the features after Embedding, that is, calculating the mean and variance of all inputs for each input respectively, and then performing standardization. This method performs a weighting operation on each feature, which can strengthen the role of effective features and weaken the interference of noise on the model

[0062] The feed-forward network includes a hidden layer network Hidden Layer and a normalization network LN_HID, and the input is V MaskedEMB , for example, setting the number of neurons to 256 and the activation function to Relu. The output of the feed-forward layer, which is also the output of a mask_block, is:

[0063] V output = Relu(LN(W i V maskedEMB ))). (3)

[0064] Concatenate the fused features output by k feature crossing modules to obtain a k-dimensional output feature, for example, expressed in the following form:

[0065] V Mask_block = concat(V output_1 , V output_2 , …, V output_k ). (4)

[0066] According to another embodiment of the present disclosure, the multi-task learning model 300 includes a multi-layer feature crossing module, which mainly constructs crossing features based on the element-wise product method. In the process of constructing crossing features, in the first layer, features in different feature domains are respectively input into corresponding feature crossing modules for learning; after the features in each feature domain are respectively learned and fused within each feature crossing module, they are all input into the second-layer feature crossing module for inter-type feature crossing learning. The functions of this layer include further learning the crossing information between feature domains, being able to more finely fuse global context information, enhancing effective features and weakening noise.

[0067] The above training sample features include: a total of m features in k feature domains corresponding to a user, and after corresponding feature embedding processing, m embedded features are obtained, where m≥2 and is an integer, and k≥2 and is an integer; the total number of features in the i-th feature domain among the above k feature domains is mi, and the value of i ranges from 1 to k. Refer to Figure 3B As shown, the above feature crossing layer 320 is a two-layer structure. The first layer includes k feature crossing modules 321 arranged in parallel, which are respectively represented as: mask_block 11~mask_block 1k, and the second layer includes r feature crossing modules 322 arranged in parallel, which are respectively represented as: mask_block 21~mask_block 2r, where r≥2 and is an integer. Specifically, the number of r is a hyperparameter, and the optimal hyperparameter r corresponding structure can be determined by evaluating the final model effect by selecting different hyperparameters as the finally selected result.

[0068] In some embodiments, each feature crossing module 321 in the first layer includes: a first feature guiding network, a first normalization network, and a first feed-forward network; each feature crossing module 322 in the second layer includes: a second feature guiding network, a second normalization network, and a second feed-forward network. In other embodiments, the feature crossing module in the second layer can also adopt the structure of a fully connected layer.

[0069] Refer to Figure 3BAs shown, in the i-th feature crossing module of the first layer, the first feature guiding network includes: a first integration layer and a first projection layer; the first integration layer is used to perform dimension expansion processing on the input mi embedded features and output a third feature, where the value of i ranges from 1 to k; the first projection layer is used to perform dimension reduction processing on the third feature and output a fourth feature; the first normalization network is used to perform normalization processing on the input mi embedded features and output a first normalized feature; perform an element-wise product operation on the third feature and the first normalized feature to obtain a first cross output result; the first feed-forward network is used to process the first cross output result and output a first fusion feature corresponding to the mi embedded features, and the first fusion feature is used as the output of the i-th feature crossing module; splice the first fusion features output by the k feature crossing modules of the first layer to obtain the first layer output feature, and the first layer output feature is used as the input to the feature crossing module of the second layer. Refer to Figure 3B As shown, in each feature crossing module of the second layer, the second feature guiding network includes a second integration layer and a second projection layer; the second integration layer is used to perform dimension expansion processing on the input k first layer output features to obtain a fifth feature; the second projection layer is used to perform dimension reduction processing on the fifth feature and output a sixth feature; the dimensions of the second integration layers in the r feature crossing modules of the second layer for dimension expansion are different; the second normalization network is used to perform normalization processing on the input k first layer output features and output a second normalized feature; perform an element-wise product operation on the sixth feature and the second normalized feature to obtain a second cross output result; the second feed-forward network is used to process the second cross output result and output a second fusion feature; splice the second fusion features output by the r feature crossing modules of the second layer to obtain the output feature of the feature crossing layer.

[0070] The processing details and examples of the feature crossing modules of the first layer and the second layer can both refer to the processing process of the feature crossing module in the previous Figure 3A shown, and will not be elaborated here.

[0071] According to an embodiment of the present disclosure, the multi-task learning layer is one of the following structures: a multi-gate mixture-of-experts network MMoE (Multi-gate Mixture-of-Experts) structure, a progressive layered extraction PLE (Progressive Layered Extraction) structure.

[0072] For example, refer to Figure 3A and Figure 3BAs shown, the above-mentioned multi-gate mixture-of-experts network structure (MMoE) includes: s expert networks (Expert), where s ≥ 2 and s is an integer, and the network parameters of each expert network are independent of each other; each expert network is used to perform network learning processing on the above output features to obtain corresponding processing results; n gating networks corresponding to n above-mentioned target scenarios, and each of the above gating networks contains s gate structures (gate) corresponding to s expert networks, and the above gate structures are used to output the selection probabilities for the corresponding expert networks; wherein, in each target scenario, the weighted sum of the processing results of the s expert networks and the corresponding selection probabilities is output.

[0073] The expert network includes multiple fully connected layers, and the number of units in the hidden layer is set to 64 (the parameters can all be adjusted according to the actual situation). The parameters between each expert network are not shared, so the outputs obtained by each expert network according to the input training are also different. In addition, the number of tasks in the multi-task learning layer is set to be equal to the number n of target scenarios. Each gate structure includes a fully connected layer, and the output is the probability of each expert network being selected. The input of the multi-task learning layer is the output V of the feature cross module mask_block Mask_block , after being processed by the multi-task learning layer, the results of each expert under different target scenarios are weighted and summed with different selection probabilities and then input into the tower layer of each task.

[0074] The fourth layer is the tower layer 340, which is connected to the multi-task learning layer 330. The input of each tower layer is the result of the weighted sum of the output of each expert and the gate of the corresponding task in the third layer (multi-task learning layer 330).

[0075] During the training process of the multi-task learning model 300, the feature embedding layer 310, the feature cross layer 320, the multi-task learning layer 330, multiple tower layers corresponding to multiple target scenarios 340, and multiple output layers 350 are trained simultaneously. Among them, for the training of the feature embedding layer 310, the dimension embeding_size after embedding the discrete variable is set to 16 (an example of a preset dimension, which can be adjusted according to the actual situation), and the batch_size is set to 512. The input in the Instance-Guided part of each mask_block is the feature after embedding, with a length of d. The number of neurons in the first fully connected layer (an example of an integration layer) is d*2 (this layer magnifies the input dimension to better capture global context information, and the magnification factor can be adjusted, with the magnification set to 2 times). The number of neurons in the second fully connected layer (an example of a projection layer) is d; the loss function for each fully connected layer is set to the Relu function f(x) = max{0, x}, making the network calculation speed and convergence speed faster; the drop_out parameter is set to 0.5, and the regularization parameter is set to 0.1. The final output after two fully connected layers (an example of a feature-guided network) is V mask = max{0, W d2 (W d1 V emb + b d1 ) + b d2}}, where W d1 and W d2 respectively represent the weight parameters of the two fully connected layers, and b d1 and b d2 respectively represent the biases of the two fully connected layers. At the same time, a LayerNormalization normalization layer is used to normalize the feature after embedding. An element-wise product operation is performed on the output of the normalization layer and the feature-guided network to obtain the cross-output result. The final cross-output result is V MaskedEMB = V mask * LN(V emb ). The number of neurons in the fully connected layer in the feed-forward layer of each mask_block layer is set to 256, and the input is V MaskedEMB . The input of the LayerNormalization layer is the output of the previous fully connected layer, and the activation function is set to the Relu function. The final function of each mask_block is V output = max{0, LN(W i V maskedEMB )}. The output of the entire mask_block layer is composed of the outputs of each mask_block concatenated together.

[0076] For each expert network, the weights and biases are initialized using a uniform distribution initializer, and L2 regularization is set to prevent overfitting. The activation function of each expert network is set to the Relu function f(x) = max{0, x}, which makes the network calculation speed and convergence speed faster. Finally, the output of each Expert is f i (x) = max{0, w i x + b}, where i = 1, 2, 3, …, s, representing the i-th expert feedforward expert network, w i is the weight of the i-th expert network, and b is the bias of the expert network.

[0077] For each gate, the activation function is set to the Softmax function:

[0078]

[0079] x is the output of the previous hidden layer of the activation function, with a dimension of s.

[0080] This activation function finally transforms the output of the previous layer into a number y between 0 and 1 i , and this result represents the probability of the current gate for the output result of each expert.

[0081] Therefore, the probabilities of each gating network gate for the results of each expert can be expressed as:

[0082] g n (x) i = softmax(w gn x + b); (6)

[0083] where n represents the n-th task, and i represents the i-th expert network; w gn is the weight of the n-th gate, and b is the bias of each gate.

[0084] Finally, the output result of the MMoE layer for each task (corresponding to the target scenario) is:

[0085]

[0086] Equation (7) represents the weighted sum of the output results of each expert and the selection probability of each gate for it; where, n represents the n-th task, and i = 1, 2, 3…s, representing the i-th expert network.

[0087] The initial weights and biases during tower layer training are determined using a normal distribution initializer, and the activation function is set to the Relu function f(x) = max{0, x}. Finally, the output of each tower layer is:

[0088] h n = max{0, wf n (x) + b}, (8)

[0089] where n represents the nth task, w is the weight of this layer, and b is the bias.

[0090] The input of each output layer is the output result of the corresponding tower layer, and the sigmoid activation function is used to output the final result. The final output result is:

[0091]

[0092] where n is the number of neurons in the last Dense layer (specifically HiddenLayer) of the output layer. The loss function is binary cross entropy:

[0093]

[0094] where N is the sample size of the training set, y i is the true value of the target value of the ith sample (using the training label), is the predicted value of the target value of the ith sample. During training, each sample of each task is identified through the corresponding indicator of each task, so as to calculate the loss during the training process of each task. The optimizer of the model uses the Adam optimizer, and the number of training epochs is set to 10. In addition, optimizers such as Stochastic Gradient Descent (SGD) and Momentum can also be used to optimize the model.

[0095] Based on the framework of the multi-task credit granting model with element-wise product feature crossing, the advantages of multi-task learning are used to solve the reusability problem of current traditional risk control models in multi-scenario and multi-label problems, realizing the situation that one model can cover customer groups or labels in multiple scenarios at the same time; at the same time, in the network, the features are weighted based on the element_wise_product method, large weights are assigned to important features to increase the role of important features, small weights are assigned to weak features, reducing the influence of weak features and noise, and increasing the fitting ability of the model.

[0096] After the model is built, the model performance can be evaluated using the auc (Area Under Curve, a common evaluation metric for classification models, whose value generally ranges from 0.5 to 1, and the larger the value, the better the model performance) and ks (Kolmogorov-Smirnov, a commonly used evaluation metric for measuring the prediction ability of a model, mainly used for binary classification problems, and the larger the value, the better the model performance) metrics. The larger these metrics are, the better the model performance.

[0097] In the embodiment including steps S210 to S230, by performing feature extraction and feature preprocessing on historical data related to user credit, training sample features are obtained; the preprocessing process can improve the data quality of the training sample features, which helps to improve the training efficiency of the multi-task learning model and the stability and accuracy of the multi-scenario credit risk control model; by fusing multi-task learning and feature crossing, feature crossing processing is performed on the input to obtain output features fused with global context information, and default prediction processing is performed under multiple target scenarios based on the above output features to obtain the corresponding default prediction probabilities under each target scenario. Feature crossing processing can not only obtain the interaction information between global features, but also assign higher weights to important features, making important features more prominent and reducing the influence of weak features and noise. Multi-task learning can cover multiple target scenarios simultaneously, effectively avoiding the cumbersome process of modeling and parameter adjustment for multiple single scenarios, and realizing risk control prediction for multiple target scenarios in the same model and effectively improving the prediction accuracy under multiple target scenarios.

[0098] The second exemplary embodiment of the present disclosure provides a method for multi-scenario user credit risk control processing.

[0099] Figure 4 Schematically shows a flowchart of a method for multi-scenario user credit risk control processing according to an embodiment of the present disclosure.

[0100] Refer to Figure 4 As shown, the method for multi-scenario user credit risk control processing provided by the embodiment of the present disclosure includes the following steps: S410, S420, and S430.

[0101] In step S410, credit-related data of the user to be evaluated is obtained.

[0102] In step S420, feature extraction and feature preprocessing are performed on the above credit-related data to obtain user features.

[0103] Details of feature extraction and feature preprocessing in this embodiment can refer to the relevant description of the first embodiment and will not be elaborated here.

[0104] In step S430, the above user characteristics are input into a pre-constructed multi-scenario credit risk control model, and the default prediction probabilities of the above user to be evaluated in multiple target scenarios are output. Among them, the above multi-scenario credit risk control model is constructed by using the method for constructing a multi-scenario credit risk control model as described above.

[0105] The third exemplary embodiment of the present disclosure provides a device for constructing a multi-scenario credit risk control model.

[0106] Figure 5 Schematically shows a structural block diagram of a device for constructing a multi-scenario credit risk control model according to an embodiment of the present disclosure.

[0107] Referring to Figure 5 As shown, the device 500 for constructing a multi-scenario credit risk control model in this embodiment includes: a first data acquisition module 501, a first feature extraction and preprocessing module 502, and a training module 503.

[0108] The above first data acquisition module 501 is used to acquire training data, and the training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the credit statuses are used to indicate whether a user defaults in each target scenario.

[0109] The above first feature extraction and preprocessing module 502 is used to perform feature extraction and feature preprocessing on the above historical data related to user credit to obtain training sample features.

[0110] The above training module 503 is used to input the above training sample features into a multi-task learning model to be trained. The output of the multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training labels are the credit statuses in the corresponding target scenarios; the trained multi-task learning model is used as the multi-scenario credit risk control model. Among them, the above multi-task learning model is used for: performing feature cross-processing on the input to obtain output features fused with global context information; and performing default prediction processing in multiple target scenarios based on the above output features to obtain the corresponding default prediction probabilities for each target scenario.

[0111] For more details of this embodiment, reference can be made to the relevant description of the first embodiment, which will not be elaborated here.

[0112] The fourth exemplary embodiment of the present disclosure provides a device for multi-scenario user credit risk control processing.

[0113] Figure 6 Schematically shows a structural block diagram of a device for multi-scenario user credit risk control processing according to an embodiment of the present disclosure.

[0114] Referring to Figure 6As shown in the figure, the multi-scenario user credit risk control processing device 600 of this embodiment includes: a second data acquisition module 601, a second feature extraction and preprocessing module 602, and a processing module 603.

[0115] The above-mentioned second data acquisition module 601 is used to acquire credit-related data of the user to be evaluated. The above-mentioned second feature extraction and preprocessing module 602 is used to perform feature extraction and feature preprocessing on the above-mentioned credit-related data to obtain user features.

[0116] The above-mentioned processing module 603 is used to input the above-mentioned user features into a pre-constructed multi-scenario credit risk control model, and output the default prediction probability of the above-mentioned user to be evaluated in multiple target scenarios; among them, the above-mentioned multi-scenario credit risk control model is constructed by using the method for constructing a multi-scenario credit risk control model provided in the first embodiment.

[0117] For more details of this embodiment, reference can be made to the relevant description of the first embodiment, which will not be elaborated here.

[0118] Any of the functional modules included in the above-mentioned device 500 or device 600 can be combined into one module for implementation, or any one of them can be split into multiple modules. Or, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. At least one of the functional modules included in device 500 or device 600 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by any other reasonable means such as integrating or packaging circuits, etc., in hardware or firmware, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of several of them. Or, at least one of the functional modules included in device 500 or device 600 can be at least partially implemented as a computer program module, and when the computer program module runs, it can execute the corresponding functions.

[0119] The fifth exemplary embodiment of the present disclosure provides an electronic device.

[0120] Figure 7 The structural block diagram of the electronic device provided by the embodiment of the present disclosure is schematically shown.

[0121] Refer to Figure 7As shown in the figure, the electronic device 700 provided by the embodiments of the present disclosure includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704. Among them, the processor 701, the communication interface 702, and the memory 703 complete communication with each other through the communication bus 704; the memory 703 is used to store a computer program; the processor 701 is used to implement the method for constructing a multi-scenario credit risk control model or the method for multi-scenario user credit risk control processing as described above when executing the program stored on the memory.

[0122] The sixth exemplary embodiment of the present disclosure further provides a computer-readable storage medium. A computer program is stored on the above computer-readable storage medium, and when the computer program is executed by a processor, the method for constructing a multi-scenario credit risk control model or the method for multi-scenario user credit risk control processing as described above is implemented.

[0123] The computer-readable storage medium may be included in the device or apparatus described in the above embodiments; it may also exist separately without being assembled into the device or apparatus. The above computer-readable storage medium carries one or more programs, and when the one or more programs are executed, the method according to the embodiments of the present disclosure is implemented.

[0124] According to the embodiments of the present disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: portable computer disks, hard disks, random access memories (RAMs), read-only memories (ROMs), erasable programmable read-only memories (EPROMs or flash memories), portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program may be used by or in combination with an instruction execution system, device, or device.

[0125] It should be noted that in the technical solutions provided by the embodiments of the present disclosure, in terms of the collection, collection, update, analysis, processing, use, transmission, storage, etc. of user personal information, they all comply with the provisions of relevant laws and regulations, are used for legal purposes, and do not violate public order and good customs. Necessary measures are taken for user personal information to prevent illegal access to user personal information data, and to maintain the security of user personal information, network security, and national security.

[0126] It should be noted that in this text, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0127] The above are only specific embodiments of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for constructing a multi-scenario credit risk control model, characterized in that, it includes: Obtain training data, where the training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the credit status is used to indicate whether the user defaults in each target scenario; Perform feature extraction and feature preprocessing on the historical data related to user credit to obtain training sample features; Input the training sample features into a multi-task learning model to be trained. The output of the multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training labels are the credit statuses in the corresponding target scenarios; The trained multi-task learning model is used as a multi-scenario credit risk control model; Among them, the multi-task learning model is used for: performing feature cross-processing on the input to obtain output features fused with global context information; and performing default prediction processing in multiple target scenarios based on the output features to obtain the corresponding default prediction probabilities for each target scenario.

2. The method according to claim 1, characterized in that, The multi-task learning model includes: A feature embedding layer for performing feature embedding processing on the training sample features to obtain embedding features of a preset dimension; A feature cross layer for performing feature cross-processing on the embedding features to obtain output features fused with global context information; A multi-task learning layer for performing default prediction learning in multiple target scenarios on the output features to obtain the learning results in each target scenario; Multiple tower layers corresponding to multiple target scenarios, each tower layer is a fully connected layer structure, and the input of each tower layer is the learning result in the corresponding target scenario; Multiple output layers corresponding to multiple target scenarios, each output layer is a fully connected layer structure, the input of each output layer is the output result of the corresponding tower layer, and each output layer outputs the default prediction probability in the corresponding target scenario.

3. The method according to claim 2, characterized in that, The training sample features include: m features corresponding to the user, and after corresponding feature embedding processing, m embedding features are obtained, where m≥2 and is an integer; The feature cross layer includes k feature cross modules arranged in parallel, and each feature cross module includes: a feature guiding network, a normalization network, and a feed-forward network, where k≥2 and is an integer; The feature guiding network includes: an integration layer and a projection layer. The integration layer is used for dimension expansion processing on the input m embedding features and outputs a first feature; the projection layer is used for dimension reduction processing on the first feature and outputs a second feature; The dimensions of dimension expansion by the integration layers in the k feature cross modules are different; The normalization network is used for normalizing the input m embedding features and outputting normalized features; Perform an element-wise product operation on the second feature and the normalized feature to obtain a cross output result; The feed-forward network is used for processing the cross output result and outputting the fused features fused with global context information corresponding to each dimension expansion dimension, and the fused features are used as the output of each feature cross module; Concatenate the fused features output by the k feature cross modules to obtain a k-dimensional output feature.

4. The method according to claim 2, wherein, The training sample features include: a total of m features in k feature domains corresponding to a user, and after performing feature embedding processing on them, m embedded features are obtained, where m≥2 and is an integer, k≥2 and is an integer; the total number of features in the i-th feature domain among the above k feature domains is m i ones, and the value range of i is 1 to k; the feature cross layer is a double-layer structure, wherein the first layer includes k feature cross modules arranged in parallel, and the second layer includes r feature cross modules arranged in parallel, r≥2 and is an integer; each feature cross module in the first layer includes: a first feature guiding network, a first normalization network, and a first feed-forward network; each feature cross module in the second layer includes: a second feature guiding network, a second normalization network, and a second feed-forward network; alternatively, the first layer includes k feature cross modules arranged in parallel, each feature cross module includes: a feature guiding network, a normalization network, and a feed-forward network, and the second layer is a fully connected layer structure; In the i-th feature cross module of the first layer, the first feature guiding network includes: a first integration layer and a first projection layer; the first integration layer is used to perform dimension expansion processing on the input m i embedded features, and output a third feature; the first projection layer is used to perform dimension reduction processing on the third feature and output a fourth feature; the first normalization network is used to perform normalization processing on the input m i embedded features, and output a first normalized feature; perform an element-wise product operation on the third feature and the first normalized feature to obtain a first cross output result; the first feed-forward network is used to process the first cross output result and output a first fusion feature corresponding to the m i embedded features, and the first fusion feature is used as the output of the i-th feature cross module; splice the first fusion features output by the k feature cross modules of the first layer to obtain a first layer output feature, and the first layer output feature is used to be input into the feature cross module of the second layer; in each feature cross module in the second layer, the second feature guiding network includes a second integration layer and a second projection layer; the second integration layer is used to perform dimension expansion processing on the input k output features of the first layer to obtain a fifth feature; the second projection layer is used to perform dimension reduction processing on the fifth feature and output a sixth feature; the dimensions of the second integration layers in the r feature cross modules in the second layer for dimension expansion are different; the second normalization network is used to perform normalization processing on the input k output features of the first layer and output a second normalized feature; an element-wise product operation is performed on the sixth feature and the second normalized feature to obtain a second cross output result; the second feed-forward network is used to process the second cross output result and output a second fusion feature; the second fusion features output by the r feature cross modules in the second layer are spliced to obtain the output feature of the feature cross layer.

5. The method according to claim 2, wherein, the multi-task learning layer is one of the following structures: a multi-gate mixture of experts network MMoE structure, a progressive layer extraction PLE structure.

6. The method according to claim 1, wherein, performing feature extraction and feature preprocessing on the user credit-related historical data to obtain training sample features, including: performing feature extraction on various types of data in the user credit-related historical data based on a feature extraction model to obtain multiple types of features; for each type of feature in the multiple types of features, calculating at least one of the following indexes of the current feature: missing rate, information value iv value, sample stability index psi value, discrimination evaluation index ks value, positive and negative sample ratio evaluation index woe value; screening the multiple types of features according to the sample screening index conditions and the indexes calculated from the multiple types of features to obtain target features for subsequent training; performing at least one of the following preprocessing operations on the target features: performing data deduplication processing, filling missing values with 0, performing one-hot encoding processing on discrete features, and performing Z-Score normalization processing on continuous features to obtain training sample features.

7. A method for multi-scenario user credit granting risk control processing, wherein, including: obtaining credit-related data of a user to be evaluated; performing feature extraction and feature preprocessing on the credit-related data to obtain user features; Input the user characteristics into a pre-constructed multi-scenario credit risk control model to output the default prediction probabilities of the user to be evaluated in multiple target scenarios; wherein, the multi-scenario credit risk control model is constructed by using the construction method described in any one of claims 1-6.

8. An apparatus for constructing a multi-scenario credit risk control model, characterized in that it includes: A first data acquisition module, configured to acquire training data, where the training data includes historical data related to user credit and credit statuses in multiple target scenarios, and the credit status is used to indicate whether the user defaults in each target scenario; A first feature extraction and preprocessing module, configured to perform feature extraction and feature preprocessing on the historical data related to user credit to obtain training sample features; A training module, configured to input the training sample features into a multi-task learning model to be trained for training. The output of the multi-task learning model is the default prediction probabilities in multiple target scenarios, and the training labels are the credit statuses in the corresponding target scenarios; the trained multi-task learning model is used as the multi-scenario credit risk control model; wherein, the multi-task learning model is configured to: perform feature cross-processing on the input to obtain output features fused with global context information; and perform default prediction processing in multiple target scenarios based on the output features to obtain the corresponding default prediction probabilities in each target scenario.

9. An apparatus for multi-scenario user credit risk control processing, characterized in that it includes: A second data acquisition module, configured to acquire data related to the credit of the user to be evaluated; A second feature extraction and preprocessing module, configured to perform feature extraction and feature preprocessing on the data related to credit to obtain user characteristics; A processing module, configured to input the user characteristics into a pre-constructed multi-scenario credit risk control model to output the default prediction probabilities of the user to be evaluated in multiple target scenarios; wherein, the multi-scenario credit risk control model is constructed by using the construction method described in any one of claims 1-6.

10. An electronic device, characterized in that it includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus; The memory is used to store a computer program; The processor is configured to implement the method described in any one of claims 1-7 when executing the program stored on the memory.

11. A computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

Images