Training method, device and electronic device for credit determination model of predetermined applicants

By building a relationship map and gradient improvement decision tree model, using the peripheral information of credit reporting households to make credit judgments, the problem of insufficient credit reporting households in the financial model is solved, and the accuracy and efficiency of credit evaluation is improved.

CN114331665BActive Publication Date: 2024-11-19ZHONGKE JUXIN INFORMATION TECH BEIJING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111333645.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-11
Publication Date
2024-11-19
Estimated Expiration
2041-11-11

AI Technical Summary

Technical Problem

The existing financial models have insufficient utilization of credit-related account data in credit risk assessment, resulting in a decrease in model effectiveness and the inability to effectively evaluate future overdue risks.

Method used

By constructing a relationship map, credit judgment is made based on the information and behavior of surrounding people, individual characteristics, graph correlation characteristics and network characteristics are extracted, and credit judgment is made using a gradient improvement decision tree model to improve the accuracy of credit judgment.

Benefits of technology

It effectively improves the accuracy of credit judgment for credit-free account, expands the scope of data use, and improves the ability of financial institutions to assess future overdue risks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114331665B_ABST
    Figure CN114331665B_ABST
Patent Text Reader

Abstract

The present application relates to a training method, device and electronic device for a credit determination model for a predetermined applicant. The training method for a credit determination model for a predetermined applicant includes: obtaining relevant data of the predetermined applicant; obtaining a relationship map for the predetermined applicant; performing sample screening based on the relationship map to obtain sample data for model establishment; extracting features based on the sample data for model establishment; and using the extracted features to train the credit determination model for the predetermined applicant. In this way, for applicants with no credit history, for example, credit determination can be performed by performing association query and feature extraction based on the relationship map, and by using the information and behavior of the applicant's surrounding people to improve the accuracy of the determination.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of financial models, and more specifically, to a training method, device and electronic device for a credit determination model for a predetermined applicant. Background Art

[0002] In recent years, Internet finance has developed rapidly, playing a positive role in promoting inclusive finance, improving the quality and efficiency of financial services, and meeting diversified investment and financing needs, showing a large market space and development potential. While injecting vitality into the development of the financial industry, Internet finance has also brought new challenges to financial management, and its rapid development has also exposed some problems and risks.

[0003] A credit-free account refers to a person who has applied for a loan or credit card but has not been approved. A blank credit record does not mean bad credit, but since financial institutions cannot grasp the credit behavior of customers, they cannot make judgments on the customer's future overdue payments, which is not conducive to the handling of business by banks or financial institutions. Generally speaking, financial institutions prefer people with good credit records, and many loan products prohibit applications from customers with a blank credit record.

[0004] As some financial institutions gradually include credit-free households in the scope of credit applications, it is crucial to effectively build a corresponding application prediction model. In the prediction of credit risk models, credit data contributes most of the prediction efficiency, but for credit-free households, the lack of credit leads to a significant reduction in the data that can be used in the model, so the efficiency of the model is also greatly reduced.

[0005] Therefore, it is desired to provide a prediction solution for households with no credit record. Summary of the invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a method, device and electronic device for training a credit judgment model for a predetermined applicant, which can improve the accuracy of the judgment by performing association query and feature extraction based on a relationship graph and using the information and behavior of the applicant's surrounding people for credit judgment for applicants with no credit history, for example.

[0007] According to one aspect of the present application, a method for training a credit determination model for a predetermined applicant is provided, comprising: obtaining relevant data of the predetermined applicant; obtaining a relationship map for the predetermined applicant; performing sample screening based on the relationship map to obtain sample data for model establishment; extracting features based on the sample data for model establishment; and using the extracted features to train the credit determination model for the predetermined applicant.

[0008] In the above-mentioned training method for the credit determination model of the predetermined applicant, obtaining the relevant data of the predetermined applicant includes data range confirmation and data preprocessing.

[0009] In the above-mentioned training method for the credit judgment model for the predetermined applicant, the construction of the relationship graph for the predetermined applicant includes: determining the data for relationship graph construction; determining the nodes and relationships in the heterogeneous graph; sorting the node attributes and relationship attributes; performing super-point screening on the heterogeneous graph; performing relationship screening on the heterogeneous graph; training the relationship weights of the heterogeneous graph; and performing community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant.

[0010] In the above-mentioned training method for the credit judgment model of the predetermined applicant, the super-point screening for the heterogeneous graph includes: defining the range of super-points, wherein the associated nodes whose in-degree number is greater than a predetermined threshold are the super-points; and, performing super-point evaluation, wherein the super-point evaluation includes at least one of the following criteria: the node is of IP type, and the average value of the median difference of the most recent login time of the associated customer nodes is greater than the threshold; the node is of address type, and the average value of the median difference of the most recent input time of the associated customer nodes is greater than the threshold; the node is of device type, and the average value of the median difference of the most recent login time of the associated customer nodes is greater than the threshold; and the node is of address type and is a non-standard location.

[0011] In the above-mentioned training method for the credit determination model of the predetermined applicant, the relationship screening for the heterogeneous graph includes: selecting a relationship type that meets at least one of a low same-label verification index, a low different-label verification index, a high connectivity coefficient index and a high average clustering coefficient index as the selected relationship.

[0012] In the above-mentioned training method for the credit determination model of the predetermined applicant, the same-label verification index is used to reflect the degree of association between the positive sample and the negative sample user nodes in the network. The calculation result is equal to the number of edges / total edges formed by the positive sample and the negative sample nodes, and is expressed as:

[0013]

[0014] In the above-mentioned training method for the credit determination model of the predetermined applicant, the different label inspection index is used to reflect the closeness between the positive sample and the negative sample nodes, and the calculation result is equal to the observed value of the edge formed by the positive sample and the negative sample nodes / the expected value of the edge formed by the positive sample and the negative sample nodes, and if the index result is less than 1, it means that the association between the positive sample and the negative sample nodes is sparse compared with the random network, and is expressed as:

[0015]

[0016] In the above-mentioned training method for the credit determination model of the predetermined applicant, the connectivity coefficient indicator is used to reflect the density of the network. The closer the value is to 1, the denser the network is, and it is expressed as:

[0017]

[0018] Where β is the number of connected branches, N i is the number of nodes in the i-th connected branch, N is the total number of nodes in the network, and L i is the average shortest path of the i-th connected branch.

[0019] In the above-mentioned training method for the credit determination model of the predetermined applicant, the average clustering coefficient index is equal to the sum of the node clustering coefficients of all nodes in the network, and the node clustering coefficient is the ratio between the actual number of edges between all neighboring nodes (excluding node i) directly connected to node i and the maximum possible number of edges between these neighboring nodes, expressed as:

[0020]

[0021] Where K represents the number of edges of all neighboring nodes of node i, and L represents the degree of node i.

[0022] In the above-mentioned training method for the credit determination model of the predetermined applicant, the training of the relationship weight of the heterogeneous graph includes: calculating the relationship weight of the heterogeneous graph based on the association property multiplied by the time attribute, expressed as:

[0023] Weight = a × e -b

[0024] Where a represents the value of the same-label inspection index in the isomorphic graph of the relationship, and b represents the time decay coefficient.

[0025] In the above-mentioned training method for the credit determination model for the predetermined applicant, community discovery is performed on the heterogeneous graph to obtain the relationship graph for the predetermined applicant, including: filtering the edges with relatively weak connections in the heterogeneous graph by edge weights; each single node in the heterogeneous graph is classified into a community, and the modularity is calculated and allocated to obtain the initial partition result; the initial partition result is compressed to generate a new graph; seed nodes in the new graph are selected, and the seed nodes meet the conditions: node degree>average degree+standard deviation; and non-seed nodes in the new graph are merged into the community of the seed nodes according to the modularity.

[0026] In the above-mentioned training method for the credit judgment model for the predetermined applicant, sample screening based on the relationship graph to obtain sample data for model establishment includes: in the relationship graph, a graph of credit-cleared samples with greater than or equal to two users and at least one with credit performance is selected as a modeling sample and a feature derivation range.

[0027] In the above-mentioned training method for the credit judgment model of the predetermined applicant, in the relationship graph, a graph with two or more users and at least one credit-clear household sample with credit performance is selected as a modeling sample and the feature derivation range includes: selecting credit-clear household samples that meet certain overdue conditions as negative samples, and selecting credit-clear household samples that have no overdue payments within a certain performance period as positive samples; other non-white household samples in the network are not used as modeling samples, but only as surrounding neighbors of the modeling samples to contribute their graph feature information to the white household samples; and, if the amount of modeling sample data is insufficient, the credit information of a part of the non-white household samples is deleted through a predetermined sampling method, and they are used as credit-clear household samples.

[0028] In the above-mentioned training method for the credit determination model of a predetermined applicant, the extracted features include individual features, graph association features and network features.

[0029] In the above-mentioned training method for the credit determination model of a predetermined applicant, the individual feature extraction performs feature extraction on the self-data information of all samples in the network, including modeling samples and non-modeling samples, and is irrelevant to the information of the user in the network.

[0030] In the above-mentioned training method for a credit determination model for a predetermined applicant, the individual characteristics include automatic behavior characteristics and attribute characteristics.

[0031] In the above-mentioned training method for the credit determination model for the predetermined applicant, the graph-associated features are individual features of other samples in the same network as the modeled sample, and are features calculated in a predetermined aggregation manner based on the network structure with the modeled sample in the graph.

[0032] In the above-mentioned training method for the credit determination model of the predetermined applicant, the network features include heterogeneous graph features and homogeneous graph features.

[0033] In the above-mentioned training method for the credit determination model for the predetermined applicant, it further comprises: using the Pearson correlation coefficient to screen the extracted features to remove irrelevant features.

[0034] In the above-mentioned training method for the credit determination model for booking applicants, the credit determination model for booking applicants is a gradient boosting decision tree model.

[0035] According to another aspect of the present application, there is provided a training device for a credit determination model for a predetermined applicant, comprising: a data acquisition unit for acquiring relevant data of the predetermined applicant; a relationship acquisition unit for acquiring a relationship map for the predetermined applicant; a sample acquisition unit for performing sample screening based on the relationship map to obtain sample data for model establishment; a feature extraction unit for extracting features based on the sample data for model establishment; and a model training unit for training the credit determination model for the predetermined applicant using the extracted features.

[0036] According to another aspect of the present application, an electronic device is provided, comprising: a processor; and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the training method for the credit determination model for a predetermined applicant as described above.

[0037] According to another aspect of the present application, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a computing device, the method for training a credit determination model for a predetermined applicant as described above is operable.

[0038] The training method, device and electronic device for the credit judgment model of a predetermined applicant provided in the present application can improve the accuracy of the judgment by conducting association queries and feature extraction based on relationship graphs for applicants with clear credit records and by using the information and behavior of people around the applicants to make credit judgments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By reading the detailed description of the preferred specific embodiments below, various other advantages and benefits of the present application will become clear to those of ordinary skill in the art. The drawings in the specification are only used for the purpose of illustrating the preferred embodiments and are not considered to be limitations of the present application. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings.

[0040] Figure 1 An exemplary flow chart of a method for training a credit determination model for a predetermined applicant according to an embodiment of the present application is illustrated;

[0041] Figure 2 A schematic diagram illustrating a homogeneous graph representing user-related relationships according to an embodiment of the present application is illustrated;

[0042] Figure 3A schematic diagram illustrating a heterogeneous graph representing relationships related to users according to an embodiment of the present application is illustrated;

[0043] Figure 4 A schematic flow chart of the construction of a relationship map in a credit determination method for a predetermined applicant according to an embodiment of the present application is illustrated;

[0044] Figure 5 A schematic diagram of an optimized large-scale network-based Louvain method according to an embodiment of the present application is illustrated;

[0045] Figure 6 A schematic diagram illustrating extraction of graph association features according to an embodiment of the present application is illustrated;

[0046] Figure 7 A schematic diagram illustrating a pseudo code of a gradient boosting decision tree model according to an embodiment of the present application

[0047] Figure 8 A block diagram of a device for training a credit determination model for a predetermined applicant according to an embodiment of the present application is illustrated;

[0048] Fig. 9 A block diagram of an electronic device according to an embodiment of the present application is illustrated. DETAILED DESCRIPTION

[0049] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0050] Exemplary Methods

[0051] Figure 1 An exemplary flow chart of a method for training a credit determination model for a predetermined applicant according to an embodiment of the present application is illustrated.

[0052] like Figure 1 As shown, the training method for the credit determination model of a predetermined applicant according to an embodiment of the present application includes the following steps.

[0053] Step S110, obtaining the relevant data of the predetermined applicant. Here, the relevant data of the predetermined applicant may be various data related to the credit judgment of the user, for example, for the credit-free household as described above, its financial asset data, SDK (software development kit) terminal operation behavior data, credit application data, etc. may be obtained, and for the non-credit-free user, in addition to obtaining the above information, its People's Bank of China credit data may also be obtained.

[0054] Specifically, in the embodiment of the present application, obtaining the relevant data of the predetermined applicant may further include two steps: data range confirmation and data preprocessing.

[0055] First, it is necessary to confirm the scope of the data source. Usually, the user information used for credit judgment comes from the personal information filled out by the user, application information, financial asset information within the bank, equipment information collected after user authorization, personal credit data, etc. In addition, it also includes overdue customer lists and blacklists within financial institutions, which are all within the scope of the data source.

[0056] Next, the acquired data needs to be preprocessed. Specifically, on the one hand, data cleaning is performed to prepare for the extraction of individual user characteristics. For example, abnormal dirty data and missing data in massive data such as user's financial asset data, SDK-side operation behavior data, credit application data, and PBOC credit data should be removed or replaced with abnormal values. On the other hand, network-related entities are normalized through data preprocessing, mainly address data. For example, address normalization processing is completed through address standardization tools to improve the accuracy of network matching.

[0057] Step S120, obtaining a relationship map for the predetermined applicant. Here, the relationship map used to represent the relationship related to the user can be pre-constructed, which mainly includes network design, weight training, heterogeneous graph construction, homogeneous graph construction, community discovery, etc. In this way, when making a credit judgment, the already constructed relationship map can be directly used. Of course, the relationship map can also be constructed when making a credit judgment.

[0058] Figure 2 FIG. 1 is a schematic diagram showing a homogeneous graph representing user-related relationships according to an embodiment of the present application. Figure 2 As shown, the nodes in the isomorphic graph are all users, which can be associated through an undirected graph or a directed graph. The relationships include various relationships between users that need to be considered, such as transfer relationships, same-device relationships, same-home address relationships, same-company address relationships, same-mobile phone number relationships, etc.

[0059] Figure 3 FIG. 1 is a schematic diagram of a heterogeneous graph representing user-related relationships according to an embodiment of the present application. Figure 3 As shown, the nodes in the heterogeneous graph may include different types of nodes, such as user nodes and associated nodes. Specifically, the user node may point to the associated node through a directed graph. The relationships between nodes include login device relationships, address relationships, login IP relationships, mobile phone number relationships, etc.

[0060] In addition, community discovery, such as the community discovery algorithm, is a graph algorithm that is used to partition based on an already constructed graph to obtain a relationship graph.

[0061] Next, the process of constructing the relationship map for the predetermined applicant will be described in detail.

[0062] Here, the relationship map for the predetermined applicants may also be referred to as a relationship network, and therefore, the process of constructing the relationship map for the predetermined applicants may also be referred to as network construction.

[0063] Figure 4 A schematic flowchart of the relationship map construction in the credit determination method for a predetermined applicant according to an embodiment of the present application is illustrated.

[0064] like Figure 4 As shown, the relationship graph construction mainly includes the following steps.

[0065] Step S121, determine the data used to build the relationship map, that is, perform data exploration. Specifically, data exploration includes statistics on the inventory data of each product, including the number of customers, the number of primary keys, the effective business volume, the number of positive and negative samples, etc.; and statistics on the status of various related dimension data, including missing rate analysis, outlier analysis, negative sample operation process characteristic induction analysis, etc.

[0066] Step S122, determine the nodes and relationships in the heterogeneous graph. That is, perform a scheme design for the heterogeneous graph. As described above, the heterogeneous graph includes nodes and relationships, where the node type is greater than 1. In an embodiment of the present application, the nodes include user nodes, address nodes, mobile phone number nodes, device nodes, and IP nodes. The address nodes include home addresses and company addresses, and the device nodes include IMEI, MAC, etc. Therefore, a variety of relationships can be defined based on the node type, including customer-operating device, customer-mobile phone number, customer-address, customer-IP, etc.

[0067] Step S123, sorting node attributes and relationship attributes. For example, node attributes and relationship attributes can be sorted according to the node types and relationship types in the heterogeneous graph, and a node data table, a relationship data table, a customer SDK operation record table, a customer account fund detail table, etc. are generated. For example, the following Table 1 shows examples of various data tables.

[0068]

Table 1

[0069]

[0070] Step 124, super-points are screened out for the heterogeneous graph. Specifically, the super-point range is first defined. In the embodiment of the present application, the associated nodes whose in-degree number is greater than a predetermined threshold are determined as super-points. In addition, the predetermined threshold can be a numerical value, such as 100, or a range, such as the in-degree number accounts for 40% of the number of network client nodes.

[0071] Then, a super-point assessment is conducted, which may include the following criteria:

[0072] (1) The node is of IP type, and the average difference between the median of the most recent login time of the associated customer nodes is greater than a threshold (e.g., 2 weeks);

[0073] (2) The node is of address type, and the average difference between the median of the most recent input time of the associated customer nodes is greater than a threshold (e.g., 2 weeks);

[0074] (3) The node is a device type, and the average difference between the median of the most recent login time of the associated customer nodes is greater than a threshold (e.g., 2 weeks);

[0075] (4) The node is of address type and is a non-standard address (e.g., lacks street / community dimension information).

[0076] Step 125, filtering the heterogeneous graph. That is, based on the SDK end operation behavior data, credit application data and other data, a large number of associations can be sorted out. In addition to the above associations, there are also associations such as email associations and GPS associations, which need to be filtered.

[0077] In the embodiment of the present application, the above associations can be screened by calculating the association relationship screening index. Here, for each type of entity, an isomorphic graph can be constructed, and the nodes on the graph are applications. If two user IDs are associated with (at least) the same entity under this type of entity, then an edge is formed. By calculating the connectivity and homogeneity index of the isomorphic graph, the connected entities are determined.

[0078] The correlation screening index includes the same-label test index, which is used to reflect the degree of correlation between the positive sample and the negative sample user nodes in the network. The calculation result is equal to the number of edges formed by the positive sample and the negative sample nodes / total number of edges. It is expressed as:

[0079]

[0080] In addition, the correlation screening index includes a different label test index, which is used to reflect the closeness between the positive sample and the negative sample nodes. The calculation result is equal to the observed value of the edge formed by the positive sample and the negative sample nodes / the expected value of the edge formed by the positive sample and the negative sample nodes. If the index result is less than 1, it means that the association between the positive sample and the negative sample nodes is sparse compared with the random network. It is expressed as:

[0081]

[0082] In addition, the correlation screening index also includes a connectivity coefficient index, which is used to reflect the density of the network. The closer the value is to 1, the denser the network is, and a dense network is more conducive to the propagation of negative labels. It can be expressed as:

[0083]

[0084] Where β is the number of connected branches, N i is the number of nodes in the i-th connected branch, N is the total number of nodes in the network, and L i is the average shortest path of the i-th connected branch.

[0085] In addition, the association relationship screening index also includes the average clustering coefficient index. Here, the clustering coefficient of a node refers to the ratio between the actual number of edges between all neighboring nodes directly connected to node i (excluding node i) and the maximum possible number of edges between these neighboring nodes. The average clustering coefficient of the network is equal to the sum of the average value of the clustering coefficients of all nodes in the network. It is expressed as:

[0086]

[0087] Where K represents the number of edges of all neighboring nodes of node i, and L represents the degree of node i.

[0088] In this way, in the embodiment of the present application, a relationship type with a low same-label verification index, a low different-label verification index, a high connectivity coefficient index, and a high average clustering coefficient index can be selected.

[0089] Step S126, training the relationship weights of the heterogeneous graph. Here, the weights of the edges in the heterogeneous graph represent the criticality of the edge type. The sample data can be used to train the weights of various types of relationships and complete the attribute labeling. In the embodiment of the present application, the calculation of the relationship weights of the heterogeneous graph can be based on two parts, namely, the association property multiplied by the time attribute, expressed as:

[0090] Weight = a × e -b

[0091] Where a represents the value of the same-label inspection index in the isomorphic graph of the relationship, and b represents the time decay coefficient, which is generally set to 1.

[0092] Step S127, performing community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant. Specifically, the community discovery algorithm extracts the isomorphic graph of the association between user nodes and user nodes based on the association between user nodes and other types of nodes in the heterogeneous graph. The purpose of community discovery is to divide the nodes in the graph into different communities so that the connections within the community are tight and the connections between communities are sparse. There are many community discovery algorithms. In the embodiment of the present application, for example, the Louvain algorithm can be selected to complete the community discovery work. The Louvain algorithm is essentially a greedy algorithm, which is a community discovery algorithm based on modularity. The algorithm performs well in efficiency and effect, and can discover hierarchical community structures. Its optimization goal is to maximize the modularity of the entire community network.

[0093] Here, modularity is a metric for evaluating the quality of a community network division. Its physical meaning is the difference between the number of edges connecting nodes in the community and the number of edges in a random situation. Its value range is [-1 / 2, 1), expressed as:

[0094]

[0095]

[0096] Among them, A ij is the weight of the edge between node i and node j, k i represents the weighted degree of node i, c i represents the community where node i is located, and m represents the sum of the weights of all edges.

[0097] The formula definition of modularity can be simplified as follows:

[0098]

[0099] Among them, Σin represents the sum of the weights of the edges in community c, and Σtot represents the sum of the weights of the edges connected to the nodes in community c.

[0100] In the embodiment of the present application, community discovery using the Louvain algorithm includes the following five steps.

[0101] Step 1: Each node in the graph is considered as an independent community, so the number of communities is the same as the number of nodes;

[0102] Step 2: For each node i, try to assign node i to the community where each of its neighboring nodes is located in turn, calculate the modularity change ΔQ before and after the assignment, and record the neighboring node with the largest ΔQ. If maxΔQ>0, assign node i to the community where the neighboring node with the largest ΔQ is located, otherwise it remains unchanged.

[0103] Step 3: Repeat step 2 until the communities to which all nodes belong no longer change;

[0104] Step 4: Compress the graph, compress all nodes in the same community into a new node, convert the edge weights between nodes in the community into the ring weights of the new node, and convert the edge weights between communities into edge weights between new nodes;

[0105] Step 5: Repeat steps 1 to 4 until the modularity of the entire graph no longer changes.

[0106] However, the number of iterations of step 2 of the above Louvain algorithm is too large, resulting in a slow running time, excessive merging of large communities, and some communities being too large; in addition, there are a large number of small communities, and small communities are not merged in time in the final division result. Based on the above problems, in the embodiment of the present application, a Louvain optimization algorithm based on a large-scale network is further proposed.

[0107] Figure 5 A schematic diagram of an optimized large-scale network-based Louvain method according to an embodiment of the present application is illustrated.

[0108] like Figure 5 As shown, the Louvain method first performs graph extraction, that is, filtering the weakly connected edges in the heterogeneous graph by edge weights; then, the initial community division is performed, that is, the single nodes in the heterogeneous graph are each classified into a community, and the modularity is calculated and distributed to obtain the initial division result; then, community compression is performed, that is, the initial division result is compressed to generate a new graph; next, the seed node in the new graph is selected, and the seed node meets the conditions: node degree>average degree+standard deviation; finally, other nodes are merged, that is, the non-seed nodes in the new graph are merged into the community of the seed node according to the modularity.

[0109] Step S130, sample screening is performed based on the relationship graph to obtain sample data for model building. For example, in the relationship graph obtained as described above, a graph with a customer number of ≥2 and at least one credit-free sample with credit performance can be selected as a modeling sample and feature derivation range.

[0110] Specifically, in an embodiment of the present application, samples of credit-free households that meet certain overdue conditions can be selected as negative samples, and samples of credit-free households that have no overdue payments within a certain performance period can be selected as positive samples; other non-white household samples in the network are not used as modeling samples, but only as surrounding neighbors of the modeling samples (white household samples) to contribute their graph feature information to the white household samples; if the amount of modeling sample data is insufficient, the credit information of some non-white household samples will be deleted through a predetermined sampling method, and they will be treated as credit-free household samples.

[0111] Step S140, extracting features based on the sample data used for model building. Here, features may include individual features, graph association features, and network features. Among them, individual features are the inherent attributes of the modeling sample, such as financial asset features and SDK operation behavior features. They can be obtained by deriving features based on the customer's own attribute information through an automated feature engineering framework (such as Featuretools).

[0112] Graph association features are attribute features of other samples in the same network as the modeling sample (such as financial asset features, SDK features, PBOC credit features, etc.). They are features calculated based on the network structure aggregation of the association relationship, association weight, association distance, etc. with the modeling sample in the graph. That is, the attributes of the customers around the modeling sample are calculated in a certain aggregation method as the attribute features of the modeling sample.

[0113] In addition, network features are structural features of the associated network in which the user is located.

[0114] In the embodiment of the present application, individual feature extraction can be performed on the data information of all samples in the network, including modeling samples and non-modeling samples, and has nothing to do with the user's information in the network. In addition, according to the feature extraction method and data source, it can be divided into automatic behavior features and attribute features.

[0115] First of all, the data sources that the automatic behavioral features rely on are the user's credit information, SDK operation record table, and user's account fund statement, etc. The generated features are mainly the user's behavioral data in the financial institution. For example, in an embodiment of the present application, the automatic extraction of customer behavior features can be achieved through an automated feature engineering framework (such as Featuretools). Here, Featuretools is good at converting time and relationship data sets into feature matrices for machine learning. It is actually an automated encapsulation of common feature engineering methods, such as groupby, min, max, mean, etc. In an embodiment of the present application, Featuretools can be used to build a unique operator for customized feature generation.

[0116] In the embodiment of the present application, the entity set of Featuretools may include a user entity table, a user credit detail table, a user SDK operation record table, and a user account fund detail table, etc. The relationship set refers to the definition of the association key between entity sets. In the embodiment of the present application, the association keys of the three data tables are, for example, user IDs.

[0117] The implementation method through Python is as follows:

[0118] Step 1: Construct entities and entity sets, expressed as:

[0119] import featuretools as ft

[0120] import pandas as pd

[0121] # Load data

[0122] customer_info=pd.read_csv('data / customer_info.csv', parse_dates=['inpute_date'])

[0123] SDK_info=pd.read_csv('data / SDK_info.csv', parse_dates=['sdk_date'])

[0124] payments_info=pd.read_csv('data / payments_info.csv', parse_dates=['psyment_date'])

[0125] #Create entity

[0126] es=ft.EntitySet(id='customer_info')

[0127] #Add customer_info entity

[0128] es=es.entity_from_dataframe(entity_id='customer_info', dataframe=customer_info, index='customer_id', time_index='inpute_date')

[0129] #Add SDK_info entity

[0130] es=es.entity_from_dataframe(entity_id='SDK_info', dataframe=SDK_info, |index='customer_id', time_index='sdk_date')

[0131] #Add pyments entity

[0132] es=es.entity_from_dataframe(entity_id='payments_info', dataframe=payments_info, make_index=True, index='customer_id', time_index='payment_date')

[0133] Step 2: Add entity relationships and define primary key and foreign key associations between three types of entities, expressed as:

[0134] #customer_id associates customer_info and SDK_info entities

[0135] r_client_SDK_info=ft.Relationship(es[′customer_info′][′customer_id′], es[′SDK_info′][′customer_id′])

[0136] es=es.add_relationship(r_client_SDK_info)

[0137] #Associate customer_info and poyments_info entities through customer_id

[0138] r_payments_info=ft.Relationship(es[′customer_info′][′customer_id′], es[′payments_info′][′customer_id′])|

[0139] es=es.add_relationship(r_payments_info)

[0140] Step 3: Aggregate features and generate new features. Based on the user's credit details, SDK operation record table and user's account fund details table, according to business experience, key features obtain the user's SDK operation frequency and maximum fund transaction value, etc. At the same time, featuretools will automatically decompose the year, month, day and work purpose according to the index of the determined time to generate multi-dimensional features. Therefore, the example of building a custom operator is as follows:

[0141]

[0142] In addition, attribute features are extracted based on the user entity table, and feature types include discrete and continuous types according to data types. Specific features include gender, age level, current asset amount level, financial asset amount level, account opening time to current month, account opening time to current days, etc.

[0143] Graph association features are individual features of other samples in the same network as the modeling sample (such as financial asset features, SDK features, PBOC credit features, etc.). They are features calculated using certain aggregation methods (Min, Max, Avg, Count, Rating, etc.) based on the network structure such as the association relationship, association weight, and association distance with the modeling sample in the graph. That is, the attributes of users around the modeling sample are calculated using a certain aggregation method and used as the attribute features of the modeling sample.

[0144] Figure 6 FIG. 1 is a schematic diagram of extracting graph association features according to an embodiment of the present application. Figure 6 As shown, the user network can be first determined based on the network formation time range, and then the users surrounding the modeling sample can be determined in the user network based on the number of network layers or distances, and the individual dimensional features of the users can be extracted, and finally aggregated in a predetermined manner.

[0145] In addition, the extraction of network features is also included, and in the embodiment of the present application, the network features include heterogeneous graph features and homogeneous graph features.

[0146] Here, the heterogeneous graph features include local network risk features, which are used to describe the statistical characteristics of the neighboring nodes of the application, including neighboring node risk features, quadrilateral risk features, and local aggregation coefficient risk features, such as the number of fraudulent applications associated with the current application, the proportion of fraudulent applications in the applications associated with the current application, etc.

[0147] In addition, heterogeneous graph features also include global network risk features, which are used to describe the risk situation of the current application. After obtaining the fraud probability of each node through a graph mining algorithm (such as PageRank), the fraud probability of the current application, the maximum and average values ​​of the fraud probability of neighboring nodes (information nodes that are associated once), etc. can be calculated.

[0148] The isomorphic graph features refer to the group features of the community where the nodes in the isomorphic graph are located, which are generated based on the community discovery clustering results, as shown in Table 2 below.

[0149]

Table 2

[0150]

[0151] In addition, in the embodiment of the present application, feature screening can be performed for the extracted features. For example, the extracted features can be screened using the Pearson correlation coefficient to remove non-related features, which is expressed as:

[0152]

[0153] Step S150: using the extracted features to train the credit determination model for the predetermined applicant.

[0154] Specifically, in the embodiment of the present application, the credit determination model for the predetermined applicant can be a GBDT (Gradient Boosting Decision Tree) model, which is a machine learning model framework with very good performance and is widely used in the industry. In the embodiment of the present application, the code of the GBDT model can be based on the third-party library sklearn of python, and use the absolute value loss function to improve the robustness of the GBDT model. Here, Figure 7 A schematic diagram of a pseudo code of a gradient boosting decision tree model according to an embodiment of the present application is illustrated.

[0155] Moreover, during the model training process, in order to ensure that the trained model can perform well on new and unseen data, data segmentation can be performed to simulate new and unseen data, that is, the training data can be divided into two parts: training samples and test samples. Next, the prediction model is trained using the training set, and then the trained model is applied to the test set for prediction.

[0156] In this way, after the credit judgment model for the predetermined applicant is trained, it is possible to perform association query and feature extraction on the data of applicants with clear credit records, and to achieve prediction and judgment based on the information and behavior of people around the applicants, thereby improving the accuracy of the judgment.

[0157] Exemplary Devices

[0158] Figure 8 The figure illustrates a block diagram of a device for training a credit determination model for a predetermined applicant according to an embodiment of the present application.

[0159] like Figure 8 As shown, according to an embodiment of the present application, a training device 200 for a credit determination model for a predetermined applicant includes: a data acquisition unit 210, used to acquire relevant data of the predetermined applicant; a relationship acquisition unit 220, used to acquire a relationship map for the predetermined applicant; a sample acquisition unit 230, used to perform sample screening based on the relationship map to obtain sample data for model establishment; a feature extraction unit 240, used to extract features based on the sample data for model establishment; and a model training unit 250, used to train the credit determination model for the predetermined applicant using the extracted features.

[0160] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the data acquisition unit 210 is used for data range confirmation and data preprocessing.

[0161] In one example, in the above-mentioned training device 200 for the credit determination model for the predetermined applicant, the relationship acquisition unit 220 is used to: determine the data for constructing the relationship graph; determine the nodes and relationships in the heterogeneous graph; sort the node attributes and relationship attributes; perform super-point screening on the heterogeneous graph; perform relationship screening on the heterogeneous graph; train the relationship weights of the heterogeneous graph; and perform community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant.

[0162] In one example, in the above-mentioned training device 200 for the credit judgment model of the predetermined applicant, the relationship acquisition unit 220 performs super-point screening on the heterogeneous graph, including: defining the range of super-points, wherein the associated nodes whose in-degree number is greater than a predetermined threshold are the super-points; and, performing super-point evaluation, wherein the super-point evaluation includes at least one of the following criteria: the node is of IP type, and the average value of the median difference of the most recent login time of the associated customer nodes is greater than the threshold; the node is of address type, and the average value of the median difference of the most recent input time of the associated customer nodes is greater than the threshold; the node is of device type, and the average value of the median difference of the most recent login time of the associated customer nodes is greater than the threshold; and the node is of address type and is a non-standard location.

[0163] In one example, in the above-mentioned training device 200 for the credit judgment model of the predetermined applicant, the relationship acquisition unit 220 performs relationship screening on the heterogeneous graph, including: selecting a relationship type that meets at least one of a low same-label verification index, a low different-label verification index, a high connectivity coefficient index, and a high average clustering coefficient index as the selected relationship.

[0164] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the same-label verification index is used to reflect the degree of association between the positive sample and the negative sample user nodes in the network, and the calculation result is equal to the number of edges / total edges formed by the positive sample and the negative sample nodes, and is expressed as:

[0165]

[0166] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the different label verification index is used to reflect the closeness between the positive sample and the negative sample nodes, and the calculation result is equal to the observed value of the edge formed by the positive sample and the negative sample nodes / the expected value of the edge formed by the positive sample and the negative sample nodes, and if the index result is less than 1, it means that the association between the positive sample and the negative sample nodes is sparse compared with the random network, and is expressed as:

[0167]

[0168] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the connectivity coefficient indicator is used to reflect the density of the network. The closer the value is to 1, the denser the network is, and it is expressed as:

[0169]

[0170] Where β is the number of connected branches, N i is the number of nodes in the i-th connected branch, N is the total number of nodes in the network, and L i is the average shortest path of the i-th connected branch.

[0171] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the average clustering coefficient indicator is equal to the sum of the node clustering coefficients of all nodes in the network, and the node clustering coefficient is the ratio between the actual number of edges between all neighboring nodes (excluding node i) directly connected to node i and the maximum possible number of edges between these neighboring nodes, expressed as:

[0172]

[0173] Where K represents the number of edges of all neighboring nodes of node i, and L represents the degree of node i.

[0174] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the training of the relationship weight of the heterogeneous graph by the relationship acquisition unit 220 includes: calculating the relationship weight of the heterogeneous graph based on the association property multiplied by the time attribute, expressed as:

[0175] Weight = a × e -b

[0176] Where a represents the value of the same-label inspection index in the isomorphic graph of the relationship, and b represents the time decay coefficient.

[0177] In one example, in the above-mentioned training device 200 for the credit determination model of the predetermined applicant, the relationship acquisition unit 220 performs community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant, including: filtering the edges with relatively weak connections in the heterogeneous graph by edge weight; classifying each single node in the heterogeneous graph into a community, and calculating the modularity for allocation to obtain the initial partitioning result; compressing the initial partitioning result to generate a new graph; selecting seed nodes in the new graph, and the seed nodes meet the conditions: node degree>average degree+standard deviation; and merging non-seed nodes in the new graph into the community of the seed nodes according to the modularity.

[0178] In one example, in the above-mentioned training device 200 for the credit judgment model of a predetermined applicant, the sample acquisition unit 230 is used to: in the relationship graph, select a graph of users greater than or equal to two and at least one credit-cleared household sample with credit performance as a modeling sample and feature derivation range.

[0179] In one example, in the above-mentioned training device 200 for the credit judgment model of the predetermined applicant, the sample acquisition unit 230 selects a graph with two or more users and at least one credit-clear household sample with credit performance in the relationship graph as a modeling sample and the feature derivation range includes: selecting credit-clear household samples that meet certain overdue conditions as negative samples, and selecting credit-clear household samples that have no overdue payments within a certain performance period as positive samples; other non-white household samples in the network are not used as modeling samples, but only as surrounding neighbors of the modeling samples to contribute their graph feature information to the white household samples; and, if the amount of modeling sample data is insufficient, the credit information of a part of the non-white household samples is deleted through a predetermined sampling method, and they are used as credit-clear household samples.

[0180] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the features extracted by the feature extraction unit 240 include individual features, graph association features and network features.

[0181] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the individual feature extraction performed by the feature extraction unit 240 performs feature extraction on the own data information of all samples in the network, including modeling samples and non-modeling samples, and is irrelevant to the user's information in the network.

[0182] In one example, in the above-mentioned training device 200 for a credit determination model of a predetermined applicant, the individual characteristics include automatic behavior characteristics and attribute characteristics.

[0183] In one example, in the above-mentioned training device 200 for the credit judgment model of the predetermined applicant, the graph-associated features are individual features of other samples in the same network as the modeled sample, and are features calculated in a predetermined aggregation manner based on the network structure with the modeled sample in the graph.

[0184] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the network features include heterogeneous graph features and homogeneous graph features.

[0185] In one example, in the above-mentioned training device 200 for the credit determination model of a predetermined applicant, the feature extraction unit 240 is further used to: screen the extracted features using the Pearson correlation coefficient to remove non-relevant features.

[0186] In one example, in the above-mentioned training device 200 for the credit determination model for the predetermined applicant, the credit determination model for the predetermined applicant is a gradient boosting decision tree model.

[0187] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the above-mentioned training device 200 for the credit determination model of a predetermined applicant have been described in the above reference. Figures 1 to 7 The training method of the credit determination model for a predetermined applicant is described in detail in the accompanying drawings, and therefore, its repeated description will be omitted.

[0188] As described above, the training device 200 for the credit determination model of the predetermined applicant according to the embodiment of the present application can be implemented in various terminal devices, such as servers of various financial institutions. In one example, the training device 200 for the credit determination model of the predetermined applicant according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the training device 200 for the credit determination model of the predetermined applicant can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the training device 200 for the credit determination model of the predetermined applicant can also be one of the many hardware modules of the terminal device.

[0189] Alternatively, in another example, the training device 200 for the credit determination model for the predetermined applicant and the terminal device may also be separate devices, and the training device 200 for the credit determination model for the predetermined applicant may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0190] Exemplary Electronic Devices

[0191] Below, reference Fig. 9 To describe an electronic device according to an embodiment of the present application.

[0192] Fig. 9 A block diagram of an electronic device according to an embodiment of the present application is illustrated.

[0193] like Fig. 9 As shown, the electronic device 10 includes one or more processors 11 and a memory 12 .

[0194] The processor 11 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 10 to perform desired functions.

[0195] The memory 12 may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may run the program instructions to implement the training method for the credit determination model of the predetermined applicant and / or other desired functions of the various embodiments of the present application described above. Various contents such as training set data, test set data, etc. may also be stored in the computer-readable storage medium.

[0196] In one example, the electronic device 10 may further include: an input device 13 and an output device 14, and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0197] For example, the input device 13 may be a keyboard, a mouse, etc.

[0198] The output device 14 can output various information to the outside, such as the credit judgment result of the predetermined applicant, etc. The output device 14 can include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, and the like.

[0199] Of course, to simplify, Fig. 9 Only some of the components related to the present application in the electronic device 10 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 10 may also include any other appropriate components.

[0200] Exemplary computer program products and computer-readable storage media

[0201] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the method for training a credit determination model for a predetermined applicant according to various embodiments of the present application described in the above-mentioned "Exemplary Method" section of this specification.

[0202] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0203] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, causes the processor to execute the steps of a method for training a credit determination model for a predetermined applicant according to various embodiments of the present application described in the above “Exemplary Method” section of this specification.

[0204] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0205] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0206] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0207] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0208] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0209] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for training a credit determination model for a predetermined applicant, characterized in that: include: Obtaining relevant data of the predetermined applicant; obtaining a relationship map for the predetermined applicant; Perform sample screening based on the relationship map to obtain sample data for model building; Extracting features based on the sample data for model building; The features include individual features, graph association features and network features; as well as Using the extracted features to train the credit determination model for the predetermined applicant; Among them, the sample screening based on the relationship graph to obtain sample data for model establishment includes: in the relationship graph, selecting a graph of a sample of a credit-free household with two or more users and at least one sample with credit performance as a modeling sample and a feature derivation range; specifically including: selecting a sample of a credit-free household that meets the overdue conditions as a negative sample, and selecting a sample of a credit-free household that has no overdue payment within a certain period of performance as a positive sample; other non-white household samples in the network are not used as modeling samples, but only as the surrounding neighbors of the modeling samples to contribute their graph feature information to the white household samples; and, if the amount of modeling sample data is insufficient, deleting the credit information of a part of the non-white household samples by a predetermined sampling method, and using them as credit-free household samples; Wherein, the individual characteristics are the attributes of the modeling sample itself; The graph-related features are individual features of other samples in the same network as the modeled sample, and are features calculated in a predetermined aggregation manner based on the network structure of the modeled sample in the graph; The network features include heterogeneous graph features and homogeneous graph features. The heterogeneous graph features include local network risk features, which are used to describe the statistical characteristics of the application's neighbor nodes, including the number of fraudulent applications associated with the current application; the heterogeneous graph features also include global network risk features, which are used to describe the risk situation of the current application, including the fraud probability of the current application.

2. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: Acquiring the relevant data of the predetermined applicant includes data range confirmation and data preprocessing.

3. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: The construction of the relationship map for the predetermined applicant includes: Determine the data used for relationship map construction; Identify nodes and relationships in a heterogeneous graph; Arrange node attributes and relationship attributes; Performing super-point screening on the heterogeneous graph; Performing relationship screening on the heterogeneous graph; Training the relation weights of the heterogeneous graph; and Performing community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant.

4. The method for training a credit determination model for a predetermined applicant as claimed in claim 3, characterized in that: The super-point elimination for the heterogeneous graph includes: Defining a super-point range, wherein the associated nodes with the number of in-degrees greater than a predetermined threshold are the super-points; and, Perform a superpoint evaluation, the superpoint evaluation including at least one of the following criteria: The node is of IP type, and the average difference between the median of the most recent login time of the associated customer nodes is greater than the threshold; The node is of address type, and the average difference of the median of the most recent input time of the associated customer node is greater than the threshold; The node is a device type, and the average of the median difference of the most recent login time of the associated customer nodes is greater than the threshold; and, The node is of address type and is non-standard.

5. The method for training a credit determination model for a predetermined applicant as claimed in claim 3, characterized in that: Relation screening for the heterogeneous graph includes: A relationship type that meets at least one of a low same-label test index, a low different-label test index, a high connectivity coefficient index, and a high average clustering coefficient index is selected as the selected relationship.

6. The method for training a credit determination model for a predetermined applicant as claimed in claim 5, characterized in that: The same-label verification index is used to reflect the degree of association between the positive sample and the negative sample user nodes in the network. The calculation result is equal to the number of edges / total edges formed by the positive sample and the negative sample nodes, and is expressed as: 。 7. The method for training a credit determination model for a predetermined applicant as claimed in claim 5, characterized in that: The different label test index is used to reflect the closeness between the positive sample and the negative sample nodes. The calculation result is equal to the observed value of the edge formed by the positive sample and the negative sample nodes / the expected value of the edge formed by the positive sample and the negative sample nodes. If the index result is less than 1, it means that the association between the positive sample and the negative sample nodes is sparse compared with the random network, and is expressed as: 。 8. The method for training a credit determination model for a predetermined applicant as claimed in claim 5, characterized in that: The connectivity coefficient index is used to reflect the density of the network. The closer the value is to 1, the denser the network is, and it is expressed as: ; Where β is the number of connected branches, is the number of nodes in the i-th connected branch, N is the total number of nodes in the network, and is the average shortest path of the i-th connected branch.

9. The method for training a credit determination model for a predetermined applicant as claimed in claim 5, characterized in that: The average clustering coefficient index is equal to the sum of the node clustering coefficients of all nodes in the network, and the node clustering coefficient is the ratio between the actual number of edges between all neighboring nodes directly connected to node i and the maximum possible number of edges between these neighboring nodes, expressed as: ; Where K represents the number of edges of all neighboring nodes of node i, and L represents the degree of node i.

10. The method for training a credit determination model for a predetermined applicant as claimed in claim 3, characterized in that: The training of the relation weights of the heterogeneous graph includes: The relationship weight of the heterogeneous graph is calculated based on the association property multiplied by the time attribute, which is expressed as: ; Where a represents the value of the same-label inspection index in the isomorphic graph of the relationship, and b represents the time decay coefficient.

11. The method for training a credit determination model for a predetermined applicant as claimed in claim 3, characterized in that: Performing community discovery on the heterogeneous graph to obtain the relationship graph for the predetermined applicant includes: Filtering the weakly connected edges in the heterogeneous graph by edge weights; The single nodes in the heterogeneous graph are respectively classified into a community, and the modularity is calculated and allocated to obtain the initial partitioning result; Compressing the initial partitioning result to generate a new graph; Selecting a seed node in the new graph, wherein the seed node meets the condition: node degree > average degree + standard deviation; and, The non-seed nodes in the new graph are merged into the community of the seed nodes according to the modularity.

12. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: The individual feature extraction is performed on the data information of all samples in the network, including modeling samples and non-modeling samples, and is irrelevant to the information of the user in the network.

13. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: The individual characteristics include automatic behavior characteristics and attribute characteristics; wherein the automatic behavior characteristics are the behavior data of the user in the financial institution, and the attribute characteristics include gender and age level.

14. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: Further including: The extracted features were screened using the Pearson correlation coefficient to remove irrelevant features.

15. The method for training a credit determination model for a predetermined applicant as claimed in claim 1, characterized in that: The credit determination model used for booking applicants is a gradient boosting decision tree model.

16. A training device for a credit determination model for a predetermined applicant, characterized in that: include: A data acquisition unit, used to acquire relevant data of the predetermined applicant; A relationship acquisition unit, used to acquire a relationship map for the predetermined applicant; A sample acquisition unit, used for performing sample screening based on the relationship map to obtain sample data for model establishment; A feature extraction unit, used to extract features based on the sample data used for model building, wherein the features include individual features, graph association features and network features; as well as A model training unit, used for training the credit determination model for the predetermined applicant using the extracted features; Among them, the sample screening based on the relationship graph to obtain sample data for model establishment includes: in the relationship graph, selecting a graph of a sample of a credit-free household with two or more users and at least one sample with credit performance as a modeling sample and a feature derivation range; specifically including: selecting a sample of a credit-free household that meets the overdue conditions as a negative sample, and selecting a sample of a credit-free household that has no overdue payment within a certain period of performance as a positive sample; other non-white household samples in the network are not used as modeling samples, but only as the surrounding neighbors of the modeling samples to contribute their graph feature information to the white household samples; and, if the amount of modeling sample data is insufficient, deleting the credit information of a part of the non-white household samples by a predetermined sampling method, and using them as credit-free household samples; Wherein, the individual characteristics are the attributes of the modeling sample itself; The graph-related features are individual features of other samples in the same network as the modeled sample, and are features calculated in a predetermined aggregation manner based on the network structure of the modeled sample in the graph; The network features include heterogeneous graph features and homogeneous graph features. The heterogeneous graph features include local network risk features, which are used to describe the statistical characteristics of the application's neighbor nodes, including the number of fraudulent applications associated with the current application; the heterogeneous graph features also include global network risk features, which are used to describe the risk situation of the current application, including the fraud probability of the current application.

17. An electronic device, characterized in that: include: processor; as well as A memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the training method for the credit judgment model of a predetermined applicant as claimed in any one of claims 1 to 15.