Abnormal repayment prediction method based on multi-task competition risk model

By using a multi-task competitive risk model, combined with multi-task learning and survival analysis, the problem of insufficient prediction of default and prepayment in traditional credit methods is solved, and accurate prediction of loan status changes over time is achieved, thus improving the comprehensiveness and accuracy of credit risk assessment.

CN119046735APending Publication Date: 2024-11-29AGRI BANK OF CHINA CO LTD CHONGQING BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411183613.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the credit field, existing technologies and traditional methods only focus on loan default risk, ignoring prepayment phenomena. Furthermore, the accuracy of survival analysis is insufficient, making it impossible to effectively predict changes in loan status over time. This leads to deficiencies in financial institutions' management of abnormal repayments.

Method used

A multi-task competitive risk model is adopted, which combines multi-task learning and competitive survival analysis. Through feature extraction and gating networks, the risk of loan default and prepayment is predicted. By utilizing unique and shared expert group networks, task interference is avoided and the prediction accuracy is improved.

Benefits of technology

It enables accurate prediction of loan status changes over time, improves the prediction accuracy of abnormal repayment phenomena, supports financial institutions in making more precise credit decisions before and during the loan process, and enhances the comprehensiveness and accuracy of credit risk assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046735A_ABST
    Figure CN119046735A_ABST
Patent Text Reader

Abstract

The invention discloses an abnormal repayment prediction method based on a multi-task competition risk model, and the method comprises the steps: inputting original data into a feature input layer based on the multi-task competition risk model, dividing a feature set data set into a continuous data set and a discrete data set, and carrying out the corresponding processing, and obtaining a coded feature set; performing feature extraction and collection on the encoded feature set in a feature extraction layer; the unique expert network of the gating network layer obtains information of the shared expert network through the gating network, the gating network outputs a feature collection result to the next layer, and the shared expert network directly outputs an attention network output result of the layer to the next layer; and the input of the gating network layer passes through the DNN full-connection neural network of the target output layer to obtain a prediction target. According to the scheme, the comprehensiveness of credit risk assessment is improved, the diversity of credit risk assessment is enriched, and the accuracy of credit risk assessment is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of financial credit risk prediction, and in particular to an abnormal repayment prediction method based on a multi-task competitive risk model. Background Art

[0002] With the rapid development of FinTech (financial technology), the barrier to obtaining loans has been significantly lowered, making the loan process more convenient and faster. Commercial banks, leveraging advanced technologies such as data science and artificial intelligence, are able to more easily assess borrowers' credit risk, thereby reducing approval costs and risks for financial institutions. However, the development of FinTech has also brought with it a series of new challenges. FinTech-based credit products, due to their simplified loan and repayment processes, can make it easier for borrowers to deviate from their repayment plans, leading to frequent irregularities such as loan defaults and prepayments. This means that while advanced information infrastructure and diversified data collection methods have attracted more customers to commercial banks, they have also placed higher demands on their pre-, mid-, and post-loan management capabilities.

[0003] In existing research on the credit field, most studies have only focused on the risk of credit default, that is, the loss of loan principal, and less attention has been paid to other abnormal repayment phenomena, such as early repayment that will result in loss of expected profits. In addition, most previous studies have only focused on whether the loan will be overdue in the future, and less attention has been paid to when this behavior will occur in the future. In past credit research, most of the focus has been on the binary prediction of credit default, that is, using a feature set X such as the loan applicant's identity information, account information, and transaction information to predict whether the loan will default after it is lent out (Y). default , where Y default It is usually a 01 variable, and its corresponding machine learning prediction estimate is usually a default probability value in the range [0,1]. Therefore, the mathematical paradigm of this type of research is usually to find an approximate solution to obtain a function f(x) such that f(X)→Y default This type of method plays an important role in risk warning for asset management and risk assessment, but traditional methods only focus on the final status of the loan, while ignoring the time when this status may occur and the potential changes over time.

[0004] Survival analysis can well capture the changes in loan status over time. When the entire loan cycle is discretized into T = (T1, T2, T3...T N )N different and adjacent periods, then the survival analysis can be used to analyze the time series T=(T1, T2, T3……T N ) is used to model the default probability in each of the N periods, and its prediction target is λdefault (T i ), where λ default (T i ) is T i The risk function of the period represents the instantaneous probability of default at that moment. The mathematical paradigm of the research at this time is to solve a function f(x) so that f(X)→λ default (T i ), the survival analysis method can clearly characterize the changes of the final state over time, but the current survival analysis methods generally have the problem of poor accuracy.

[0005] At the same time, with the development of modern credit, the loan model is no longer limited to the traditional loan model, and new loan models such as revolving loans and on-demand loans have emerged. This means that credit institutions not only have to face the loss of loan principal due to default, but also the loss of interest due to early repayment. This means that the prediction target of credit institutions has changed from the previous single target Y default , becomes the current multi-objective, such as the dual objective Y(Y default 、Y prepayment ), therefore, it may be necessary to introduce a multi-task learning framework to rebuild the mathematical model so that the modeling goal becomes f(X)→Y(Y default 、Y prepayment ), but in actual credit scenarios, customers who repay in advance often represent customers with good credit, strong repayment level and repayment willingness, while defaulting customers represent customers with poor credit, poor repayment level or repayment willingness. Their user behaviors are quite different, which may cause mutual interference between the two prediction tasks. Summary of the Invention

[0006] Based on the above technical problems, the present invention proposes an abnormal repayment prediction method based on a multi-task competitive risk model. It targets two common abnormal repayment phenomena in the credit field: default and early repayment. Combining the characteristics of credit field data, it uses multi-task learning and competitive survival analysis, and considers the risks of early repayment and default at the same time, to characterize the risk situation of the loan from the aspects of survival time and survival probability. The specific technical solution is as follows.

[0007] The abnormal repayment prediction method based on the multi-task competing risk model includes the following steps:

[0008] Step S1: Input the original data into the feature input layer of the multi-task competing risk model, and divide the feature set dataset into a continuous data set X C and discrete data set X D And perform corresponding processing to obtain the encoded feature set X E ;

[0009] Step S2: The feature set X after encoding is extracted in the multi-task competitive risk model feature extraction layer E Perform feature extraction and aggregation;

[0010] Step S3: The unique expert network in the gated network layer of the multi-task competitive risk model obtains the information of the shared expert network through the gated network, and the gated network outputs the result of feature aggregation To the next layer, the shared expert network directly outputs the attention network output of this layer Go to the next level;

[0011] Step S4: Gating the network layer input The predicted target is obtained by passing the DNN fully connected neural network of the target output layer of the multi-task competitive risk model.

[0012] Specifically, the prediction target of step S4 is the default risk function Y of the main task default (T i ), risk function λ of early repayment prepayment (T i ) and auxiliary task prediction results in, It represents the prediction of the survival time of the loan set, and its prediction value range is [0,T max ], It indicates the estimation of whether the customer has prepayment or default status, and its predicted value range is [0,1].

[0013] Specifically, the original data of step S1 is the credit survival data set D = {X, T, Y default ,Y prepayment}; where X is the original feature set of the dataset, T is the loan survival time set, and Y default and Y prepayment is a set of credit states, each y default,i and Y prepayment,i It is a 01 variable. When the value is 1, it means that the loan has this status. When the value is 0, it means that the loan has not this status. default,i and y prepayment,i The value of cannot be 1 at the same time.

[0014] Specifically, the step S1 performs the following operations on the continuous data set X: C and discrete data set X D The processing includes:

[0015] For a continuous data set X C , use the maximum and minimum standardization to standardize its values;

[0016] For a discrete data set X D , use One-hot encoding to encode the categorical variables and obtain the encoded feature set X E .

[0017] Specifically, the feature extraction layer in step S2 includes a unique expert group network, a shared expert group network, and a customized gating network. The unique expert group network and the shared expert group network each include multiple experts set as attention networks. The calculation of the attention network includes:

[0018] Perform linear transformation, the linear transformation formula is:

[0019]

[0020] Get the query vector Q i , key vector K i Sum value vector V i , and then get the output attention score E according to the attention weight formula i , E i The expression is:

[0021] E i =Score i =Attention(Q i ,K i ) i ·V i ;

[0022] The attention weight formula is specifically:

[0023]

[0024] Among them, d k It's Q i , K i dimension.

[0025] Specifically, the unique expert group network and the shared expert group network further include: calculating all expert attention scores to obtain two unique expert groups: the default prediction expert group E D (E D,1 ,E D,2 …E D,k ) and the Prepayment Prediction Expert Group E P (E P,1 ,E P,1 …E P,m ), and two different shared expert groups E S1 (E S1,1 ,E S1,1 …E S1,n1 ) and E S2 (ES2,1 、E S2,1 …E S2,n2 ); where subscripts k, m, n1, and n2 are the numbers of experts in the expert group.

[0026] Specifically, the expert groups that ultimately participate in decision-making in the customized gating network include: the shared expert group and all expert groups E′ S1 [E D 、E P 、E S1 、E S2 ] and members E′ of the exclusive expert group that does not include other exclusive expert groups D [E D 、E S1 、E S2 ], the final output of the unique expert group is:

[0027]

[0028] Among them, W D is the initial expert group weight matrix, whose first dimension is determined by the number of expert members in the expert group.

[0029] Specifically, the decision-making group member E' of the shared expert group S1 [E D 、E P 、E S1 、E S2 The final output is calculated in the same way as the final output of the unique expert group.

[0030] Beneficial effects of the present invention: The present invention proposes an abnormal repayment prediction method based on a multi-task competitive risk model, which captures more accurate risk prediction patterns, improves the commercial banks' prediction accuracy for abnormal repayment phenomena such as defaults and early repayments, and simultaneously provides information in the time dimension to support credit institutions in making more accurate and reliable credit decisions before and during loans, which has strong theoretical and practical significance.

[0031] It has the following technical effects:

[0032] (1) Improved comprehensiveness of credit risk assessment: First, traditional risk assessment methods often only consider the risk of default, without considering the risks of other abnormal repayments. This method introduces a multi-task framework that can simultaneously model multiple abnormal repayments, thereby improving the comprehensiveness of credit risk assessment.

[0033] (2) Enriching the diversity of credit risk assessment: Traditional assessment methods often fail to fully consider time factors. However, the survival analysis introduced in this method can better model time dependence, thereby improving the accuracy of credit risk assessment. By comprehensively considering the borrower's probability of default and early repayment and the time, this method can better predict risks and help financial institutions better manage credit risks.

[0034] (3) Improved the accuracy of credit risk assessment: This method, based on the two auxiliary tasks proposed in the multi-task learning framework, can improve the predictive ability of the risk model and provide financial institutions with more accurate risk assessment results, thereby helping them make better credit decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the abnormal repayment prediction method based on the multi-task competitive risk model of the present invention;

[0036] Figure 2 This is a diagram of the architecture of a multi-task competing risk model in an embodiment of the present invention;

[0037] Figure 3 is a schematic diagram of an attention network in an embodiment of the present invention;

[0038] Figure 4 Schematic diagram of a DNN fully connected neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described with reference to the accompanying drawings.

[0040] This paper proposes an abnormal repayment prediction method based on a multi-task competitive risk model. The core modeling goal is to use a multi-task learning framework to simultaneously predict two common abnormal repayment phenomena in the credit field: default Y default and early repayment Y prepayment We conduct dual-objective modeling and introduce the theoretical framework of survival analysis to characterize the changes of loans over time. Therefore, the core prediction target of this method is the risk function λ default (T i ) and λ prepayment (T i ), ultimately, by predicting the probability of loan default and early repayment in each billing period, it helps credit institutions understand the changes in loan status over time in each period. Figure 1 As shown, the following steps are included:

[0041] Step S1: Input the original data into the feature input layer of the multi-task competing risk model, and divide the feature set dataset into a continuous data set XC and discrete data set X D And perform corresponding processing to obtain the encoded feature set X E ;

[0042] Step S2: The feature set X after encoding is extracted in the multi-task competitive risk model feature extraction layer E Perform feature extraction and aggregation;

[0043] Step S3: The unique expert network in the gated network layer of the multi-task competitive risk model obtains the information of the shared expert network through the gated network, and the gated network outputs the result of feature aggregation To the next layer, the shared expert network directly outputs the attention network output of this layer Go to the next level;

[0044] Step S4: Gating the network layer input The predicted target is obtained by passing the DNN fully connected neural network of the target output layer of the multi-task competitive risk model.

[0045] In this embodiment, the prediction target of step S4 is the default risk function λ of the main task default (T i ), risk function λ of early repayment prepayment (T i ) and auxiliary task prediction results in, It represents the prediction of the survival time of the loan set, and its prediction value range is [0,T max ], It indicates the estimation of whether the customer has prepayment or default status, and its predicted value range is [0,1].

[0046] In this embodiment, the original data of step S1 is the credit survival data set D = {X, T, Y default ,Y prepayment}; where X is the original feature set of the dataset, T is the loan survival time set, and Y default and Y prepayment is a set of credit states, each y befault,i and y prepayment,i It is a 01 variable. When the value is 1, it means that the loan has this status. When the value is 0, it means that the loan has not this status. default,i and y prepayment,i The value of cannot be 1 at the same time.

[0047] In this embodiment, step S1 processes the continuous data set X. C and discrete data set X D The processing includes:

[0048] For a continuous data set X C , use the maximum and minimum standardization to standardize its values;

[0049] For a discrete data set X D , use One-hot encoding to encode the categorical variables and obtain the encoded feature set X E .

[0050] In this embodiment, the feature extraction layer in step S2 includes a unique expert group network, a shared expert group network, and a customized gating network. The unique expert group network and the shared expert group network each include multiple experts configured as attention networks. The calculation of the attention network includes:

[0051] Perform linear transformation, the linear transformation formula is:

[0052]

[0053] Get the query vector Q i , key vector K i Sum value vector V i , and then get the output attention score E according to the attention weight formula i , E i The expression is:

[0054] E i =Score i =Attention(Q i ,K i ) i ·V i ;

[0055] The attention weight formula is specifically:

[0056]

[0057] Among them, d k It's Q i , K i dimension.

[0058] In this embodiment, the unique expert group network and the shared expert group network further include: calculating all expert attention scores to obtain two unique expert groups: the default prediction expert group E D (E D,1 ,E D,2 …E D,k ) and the Prepayment Prediction Expert Group E P (E P,1 ,E P,1 …E P,m), and two different shared expert groups E S1 (E S1,1 ,E S1,1 …E S1,n1 ) and E S2 (E S2,1 、E S2,1 …E S2,n2 ); where subscripts k, m, n1, and n2 are the numbers of experts in the expert group.

[0059] In this embodiment, the expert groups that ultimately participate in decision-making in the customized gating network include: the shared expert group and all expert groups E′ S1 [E D 、E P 、E S1 、E S2 ] and members E′ of the exclusive expert group that does not include other exclusive expert groups D [E D 、E S1 、E S2 ], the final output of the unique expert group is:

[0060]

[0061] Among them, W D is the initial expert group weight matrix, whose first dimension is determined by the number of expert members in the expert group.

[0062] In this embodiment, the decision-making group members E′ of the shared expert group S1 [E D 、E P 、E S1 、E S2 The final output is calculated in the same way as the final output of the unique expert group.

[0063] In one embodiment, the design idea of ​​the multi-task competitive risk model is to use the deep learning network that has been fully verified in accuracy in recent years as the basic network, and at the same time introduce and improve the PLE (Progressive Layered Extraction) architecture, by establishing a multi-task learning framework in which unique experts and shared experts coexist, and using a gating mechanism to adaptively combine unique experts and shared experts, so as to learn the commonalities between the early repayment prediction task and the default prediction task without losing the unique characteristics of the task, so as to avoid the phenomenon of "negative transfer" of the model's performance. At the same time, in order to further improve the predictive ability of the model, the multi-task competitive risk model of the present invention relies on the multi-task learning architecture to design two related auxiliary tasks to help the model improve its performance. The specific architecture of the model is as follows: Figure 2 shown.

[0064] The entire model is divided into four templates: the feature input layer (Input Layer), the feature extraction layer (ExtractionNetwork Layer), the gated network layer (CGC Layer) and the final target output layer (Out Layer). Based on the feature set X input to the model, the final model output will obtain the risk function λ corresponding to each period of loan default and early repayment. default (T i ),λ prepayment (T i )(main task), as well as the customer's survival time (auxiliary task 1) and whether the customer has prepayment or default (auxiliary task 2).

[0065] Feature input layer: In the feature input layer, the original input data is the credit survival data set D = {X, T, Y default ,Y prepayment}, where X is the original feature set of the dataset, T is the survival time set of the loan, and Y default and Y prepayment is the set of credit states, for each y default,i and y prepayment,i It is a 01 variable. When the value is 1, it means that the loan has this status. When the value is 0, it means that the loan has not this status. Because the two states of default and early repayment cannot appear in the same loan at the same time, y default,i and y prepayment,i The value of cannot be 1 at the same time.

[0066] The survival time set T and credit status set Y entering the input layer default and Y prepayment No further data processing will be performed at this layer, but it will be directly used to update the training parameters of the subsequent model. For the feature set X, this scheme divides the data set into a continuous data set X C and discrete data set X D In the Embedding layer, for the continuous data set X C , this scheme uses the maximum and minimum standardization to perform numerical standardization on the discrete data set X D , this paper uses One-hot encoding to encode the categorical variables, and finally obtains the encoded feature set X E .

[0067] Feature extraction layer: In the feature extraction layer, the feature set X obtained by the upper layer EThis layer is where features are fully extracted and aggregated. This layer consists of a unique expert group network (Experts Pre, Experts Default), a shared expert group network (Experts Shared), and a custom gating network (G in the layer).

[0068] For the unique and shared expert group networks, each expert group is composed of several experts. In this paper, each expert in this layer is set as Figure 3 The attention network shown in the formula (4-1) can be transformed linearly into

[0069]

[0070] Get the query vector Q i , key vector K i , value vector V i , and then use formula (4-2) to get its attention weight, where d k It's Q i , K i Dimensions:

[0071]

[0072] Finally, the attention score of the expert’s output can be obtained as shown in formula (4-3), denoted as E i :

[0073] E i =Score i =Attention(Q i ,K i ) i ·V i (4—3)

[0074] After calculating all the expert attention scores, we get two unique expert groups: default prediction expert group E D (E D,1 ,E D,2 …E D,k ), Prepayment Prediction Expert Group E P (E P,1 ,E P,1 …E P,m ) and two different shared expert groups E S1 (E S1,1 ,E S1,1 …E S1,n1 ), E S2 (E S2,1 、E S2,1 …E S2,n2 ), where subscripts k, m, n1, and n2 are the numbers of experts in the expert group.

[0075] In the gated network layer ( Figure 3 In order to avoid mutual interference in the prediction tasks between unique experts, the members of the unique expert group participating in the final decision here will not include other unique expert groups, for example, the expert decision group E′ for early repayment D , whose decision-making group members are E′ D [E D 、E S1 、E S2 ], and the final output The calculation method is shown in formula (4-4).

[0076]

[0077] Where W D is the initial expert group weight matrix, whose first dimension is determined by the number of expert members in the expert group.

[0078] For the shared expert group, since its task is to gather more information, unlike the unique expert group, the final decision-making members of the shared expert group will include all expert groups. For example, for the shared expert decision group E′ S1 , whose decision-making group members are E′ S1 [E D 、E P 、E S1 、E S2 ], its final output The calculation method is the same as formula (4-4).

[0079] Gated network layer: The main components of the gated network layer are still the unique expert group network, the shared expert group network, and the customized gated network. However, unlike the feature extraction layer of the previous layer, the unique expert network of this layer will not share the information with the shared experts after passing through the same attention network as the previous layer, in order to ensure the focus on the final prediction task. However, the unique experts can obtain the information of the shared experts through the same gating network as the upper layer to help improve the accuracy of the final prediction task. The final gated network will output the result of feature aggregation. To the next layer, for the shared expert network, the model directly outputs the attention network output of this layer without going through the gating network Go to the next floor.

[0080] Target output layer: In the target output layer, the input from the gated network layer Each of them passes through a random dropout and ReLU layer as follows Figure 4 The DNN fully connected neural network shown in the figure finally obtains the prediction target of this paper as the default risk function λ of the main task default(T i ), risk function λ of prepayment prepayment (T i ) and auxiliary task prediction results It represents the prediction of the survival time of the loan set, and its prediction value range is [0,T max ], It indicates the estimation of whether the customer has made early repayment or defaulted. The predicted value range is [0, 1].

[0081] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The abnormal repayment prediction method based on the multi-task competitive risk model is characterized by: The following steps are involved: Step S1: Input the original data into the feature input layer of the multi-task competitive risk model, and divide the feature set data set into a continuous data set X C and discrete data set X D And perform corresponding processing to obtain the encoded feature set X E ; Step S2: In the multi-task competitive risk model feature extraction layer, the encoded feature set X E Perform feature extraction and aggregation; Step S3: The unique expert network of the gated network layer of the multi-task competitive risk model obtains the information of the shared expert network through the gated network, and the gated network outputs the result of feature aggregation To the next layer, the shared expert network directly outputs the attention network output result of this layer Go to the next level; Step S4: Gating the network layer input The DNN fully connected neural network of the target output layer of the multi-task competitive risk model is used to obtain the predicted target.

2. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 1 is characterized in that: The step S4 predicts the default risk function λ of the main task default (T i ), risk function of early repayment λ prepayment (T i ) and auxiliary task prediction results in, Represents the prediction of the loan set survival time, and its prediction value range is [0, T max ], It indicates the estimation of whether the customer has prepayment or default status, and its prediction value range is [0, 1].

3. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 1 is characterized in that: The original data of step S1 is the credit survival data set D = {X, T, Y default , Y prepayment }; where X is the original feature set of the data set, T is the loan survival time set, and Y default and Y prepayment is a set of credit states, each y default,i and prepayment,i is a 01 variable. When the value is 1, it means that the loan has this status. When the value is 0, it means that the loan has not this status. default,i and prepayment,i The value of cannot be 1 at the same time.

4. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 3 is characterized in that: The step S1 performs a c and discrete data set X D The processing includes: For a continuous data set X C , use the maximum and minimum standardization to standardize its values; For a discrete data set X D , use One-hot encoding to encode the categorical variables and obtain the encoded feature set X E .

5. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 1 is characterized in that: The feature extraction layer in step S2 includes a unique expert group network, a shared expert group network and a customized gating network. The unique expert group network and the shared expert group network both include a plurality of experts set as attention networks. The calculation of the attention network includes: Perform linear transformation, the linear transformation formula is: Get the query vector Q i , key vector K i Sum value vector V i , and then get the output attention score E according to the attention weight formula i , E i The expression is: E i =Score i =Attention(Q i ,K i ) i ·V i ; The attention weight formula is specifically: Among them, d k It's Q i , K i Dimension.

6. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 5 is characterized in that: The unique expert group network and the shared expert group network also include: after calculating all expert attention scores, two unique expert groups are obtained: the default prediction expert group E D (E D,1 , E D,2 …E D,k ) and the Prepayment Prediction Expert Group E P (E P,1 , E P,1 …E P,m ), and two different shared expert groups E S1 (E S1,1, E S1,1 …E S1,n1 ) and E S2 (E S2,1 、E S2,1 …E S2,n2 ), where subscripts k, m, n1, and n2 are the numbers of experts in the expert group.

7. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 6 is characterized in that: The expert groups that ultimately participate in decision-making in the custom gating network include: the shared expert group, all expert groups E′ S1 [E D 、E P 、E S1 、E S2 ] and members E′ of the exclusive expert group that will not include other exclusive expert groups D [E D 、E S1 、E S2 ], the final output of the unique expert group is: Among them, W D is the initial expert group weight matrix, whose first dimension is determined by the number of expert members in the expert group.

8. The abnormal repayment prediction method based on the multi-task competitive risk model according to claim 6 is characterized in that: The shared expert group decision group member E′ S1 [E D 、E P 、E S1 、E S2 ]The final output is calculated in the same way as the final output of the unique expert group.