Credit risk assessment method and device, equipment and storage medium
By aggregating the features of other nodes in the subgraph into the target node in the credit risk assessment, the problem of poor interpretability of the GNN model is solved, improving the accuracy and efficiency of the credit risk assessment, and providing a more intuitive business explanation.
Patent Information
- Application Number
- CN202410994960.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-01
AI Technical Summary
In credit risk assessment, the graph neural network (GNN) model is poorly interpretable, and the explanations of existing tools at the topological level are not intuitive and friendly enough, making it difficult to meet the needs of credit business.
By selecting the target node from the graph data, dividing the graph, and aggregating the user credit characteristics of other nodes in the sub-graph onto the target node, aggregating features are obtained, which are used by machine learning models to predict credit risks, and improve the interpretability of the model and Data processing efficiency.
It realizes the extraction of aggregation features with business meaning from graph data, improves the interpretability and processing efficiency of machine learning models, reduces the amount of data, and enhances the accuracy and portability of predictions.
Smart Images

Figure CN118887005A_ABST
Abstract
Description
Technical Field
[0001] One or more embodiments of this specification relate to the field of data processing, and in particular, to a credit risk assessment method, apparatus, device, and storage medium. Background Art
[0002] Since graph data can reveal hidden connections between entities, its application in credit risk assessment can provide deeper understanding and insights, helping to make more accurate and comprehensive credit decisions.
[0003] In the credit field, graph neural networks (GNNs) are usually used to process graph data to predict users' credit risks. Since GNNs make predictions by learning to capture the structural information of the graph and the hidden representations of node features, these hidden representations are abstract, resulting in poor interpretability of GNNs. Although there are many tools on the market designed to improve the interpretability of GNN models, such as PGExpliner, GNNExpliner, GraphLIME, and GAT, these tools have made contributions in edge level, graph structure interpretation, node-level feature selection, and explaining the importance of node features through attention mechanisms. However, for the credit business, these methods are still not intuitive and friendly enough in topological level interpretation. Summary of the invention
[0004] In view of this, one or more embodiments of the present specification provide a method for improving the interpretability of graph features and the interpretability of machine learning models in processing graph data.
[0005] To achieve the above objectives, one or more embodiments of this specification provide the following technical solutions:
[0006] According to a first aspect of one or more embodiments of this specification, a credit risk assessment method is proposed, comprising:
[0007] Selecting a target node to be evaluated from multiple nodes in the graph data, wherein the characteristics of the nodes in the graph data include user credit characteristics;
[0008] Determine a plurality of subgraphs in the graph data according to the target node;
[0009] aggregating the user credit feature of the first node in the subgraph to the target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node;
[0010] A target feature is selected from the aggregated features and input into a machine learning model, where the machine learning model is used to predict the credit risk of the user.
[0011] In some embodiments, determining a plurality of subgraphs in the graph data according to the target node includes:
[0012] Determine the subgraph corresponding to the target node by adopting a first sampling method;
[0013] Among them, the first sampling method and the first aggregation function have a preset mapping relationship.
[0014] In some embodiments, the method further comprises:
[0015] Combining each sampling method of the plurality of sampling methods with each aggregation function of the plurality of aggregation functions one by one to obtain a plurality of groups of sampling aggregation pairs;
[0016] Using each group of sampling aggregation pairs to process the sample graph data respectively, to obtain sample aggregation features, wherein the sample graph data is graph data with true values of credit risk assessment results marked on sample nodes;
[0017] According to the credit risk prediction results corresponding to the sample aggregation characteristics and the corresponding true values, the mapping relationship between the sampling method and the aggregation function is determined.
[0018] In some embodiments, the user credit feature of the node includes a plurality of sub-features;
[0019] Aggregating the user credit feature of the first node in the subgraph to the target node in the subgraph to obtain the aggregated feature includes:
[0020] Aggregating each sub-feature of the first node to the target node to obtain sub-aggregated features;
[0021] The aggregate feature is obtained according to each sub-aggregate feature.
[0022] In some embodiments, the method further comprises:
[0023] Performing abnormal sample detection on the aggregated features and removing abnormal samples;
[0024] The aggregated features are updated using the eliminated results.
[0025] In some embodiments, the method further comprises:
[0026] The aggregated features are screened to obtain target features whose correlation with the indicator to be evaluated meets the set requirements.
[0027] In some embodiments, the method further comprises:
[0028] The graph data is cleaned, wherein the data cleaning includes at least one of excluding strong linear correlation features and removing hot spot data.
[0029] According to a second aspect of one or more embodiments of this specification, a credit risk assessment device is provided, comprising:
[0030] A selection unit, configured to select a target node to be evaluated from a plurality of nodes in the graph data, wherein the characteristics of the nodes in the graph data include user credit characteristics;
[0031] A determination unit, configured to determine a plurality of subgraphs in the graph data according to the target node;
[0032] an aggregation unit, configured to aggregate the user credit feature of a first node in the subgraph to a target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node;
[0033] A prediction unit is used to select target features from the aggregated features to be input into a machine learning model, where the machine learning model is used to predict the credit risk of a user.
[0034] In some embodiments, the determining unit is specifically configured to:
[0035] Determine the subgraph corresponding to the target node by adopting a first sampling method;
[0036] Among them, the first sampling method and the first aggregation function have a preset mapping relationship.
[0037] In some embodiments, the apparatus further comprises a mapping unit configured to:
[0038] Combining each sampling method of the plurality of sampling methods with each aggregation function of the plurality of aggregation functions one by one to obtain a plurality of groups of sampling aggregation pairs;
[0039] Using each group of sampling aggregation pairs to process the sample graph data respectively, to obtain sample aggregation features, wherein the sample graph data is graph data with true values of credit risk assessment results marked on sample nodes;
[0040] According to the credit risk prediction results corresponding to the sample aggregation characteristics and the corresponding true values, the mapping relationship between the sampling method and the aggregation function is determined.
[0041] In some embodiments, the user credit feature of the node includes a plurality of sub-features;
[0042] The polymeric unit is specifically used for:
[0043] Aggregating each sub-feature of the first node to the target node to obtain sub-aggregated features;
[0044] The aggregate feature is obtained according to each sub-aggregate feature.
[0045] In some embodiments, the device further comprises an abnormality elimination unit, configured to:
[0046] Performing abnormal sample detection on the aggregated features and removing abnormal samples;
[0047] The aggregated features are updated using the eliminated results.
[0048] In some embodiments, the device further comprises a screening unit for:
[0049] The aggregated features are subjected to feature screening to obtain target features whose correlation with the indicator to be evaluated meets the set requirements. In some embodiments, the device further includes a data cleaning unit for:
[0050] The graph data is cleaned, wherein the data cleaning includes at least one of excluding strong linear correlation features and removing hot spot data.
[0051] According to a third aspect of one or more embodiments of this specification, an electronic device is provided, including:
[0052] processor;
[0053] a memory for storing processor-executable instructions;
[0054] The processor implements the steps of the method proposed in the above embodiment by running the executable instructions.
[0055] According to a fourth aspect of one or more embodiments of the present specification, a computer-readable storage medium is provided, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the method provided in the above embodiment are implemented.
[0056] According to a fifth aspect of one or more embodiments of the present specification, a computer program product is proposed, comprising a computer program / instruction, which implements the steps of the method proposed in the above embodiment when executed by a processor.
[0057] The credit risk assessment method proposed in the embodiment of this specification first selects the target node to be evaluated from multiple nodes of the graph data representing the credit characteristics of the user, and determines multiple subgraphs in the graph data according to the target node, aggregates the user credit characteristics of other nodes in the subgraph other than the target node to the target node in the subgraph to obtain aggregated characteristics, and finally selects the target characteristics input to the machine learning model from the aggregated characteristics, and the machine learning model is used to predict the credit risk of the user. On the one hand, by aggregating the user credit characteristics of other nodes in the subgraph to the target node, the aggregated characteristics with business meaning are extracted from the graph data, and the credit risk of the user is predicted based on these aggregated characteristics, which improves the interpretability of the machine learning model in the processing process; on the other hand, based on the graph data, the aggregated characteristics of each target feature are extracted for prediction, which reduces the amount of data to be processed and is also conducive to improving the interpretability of the machine learning model processing process. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 It is a flowchart of a credit risk assessment method provided by an exemplary embodiment.
[0059] Figure 2 It is a flow chart of a method for determining a mapping relationship between a sampling method and an aggregation function provided by an exemplary embodiment.
[0060] Figure 3 It is a schematic diagram of obtaining sampling aggregation pairs provided by an exemplary embodiment.
[0061] Figure 4 is a flow chart of another credit risk assessment method provided by an exemplary embodiment.
[0062] Figure 5 It is a block diagram of a credit risk assessment device provided by an exemplary embodiment.
[0063] Figure 6 It is a structural schematic diagram of a device provided by an exemplary embodiment. DETAILED DESCRIPTION
[0064] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with one or more embodiments of this specification. Instead, they are merely examples of devices and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.
[0065] It should be noted that: in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this specification may be combined into a single step for description in other embodiments.
[0066] Graph data is a data structure used to represent entities and their relationships with each other. In graph data, information is organized into nodes and edges, each node represents an entity, and the edge represents a certain relationship or interaction between nodes.
[0067] In the credit field, graph data can reveal hidden connections between transactions. Therefore, its application in credit risk assessment can provide deeper understanding and insights, helping to make more accurate and comprehensive credit decisions.
[0068] In related technologies, GNN is usually used to process graph data to predict the credit risk of users. Since GNN makes predictions by learning to capture the structural information of the graph and the hidden representation of node features, these hidden representations are abstract, resulting in poor interpretability of GNN. Although there are many tools on the market designed to improve the interpretability of GNN models, such as PGExpliner, GNNExpliner, GraphLIME, and GAT, these tools have made contributions in edge level, graph structure interpretation, node-level feature selection, and explaining the importance of node features through attention mechanisms. However, for the credit business, these methods are still not intuitive and friendly enough in topological level interpretation.
[0069] In view of this, an embodiment of the present specification proposes a credit risk assessment method, which divides graph data into multiple subgraphs based on a target node, and aggregates the features of nodes other than the target node in each subgraph to the target node to obtain aggregated features for predicting the user's credit risk.
[0070] Please refer to Figure 1 , showing a credit risk assessment method proposed in an embodiment of this specification, including steps 101 to 104.
[0071] In step 101, a target node to be evaluated is selected from multiple nodes of graph data.
[0072] In the credit field, nodes in graph data can represent entities such as borrowers and financial institutions, and edges represent credit relationships and transaction relationships between entities. When nodes represent borrowers, the characteristics of nodes in graph data can include user credit characteristics, including the user's credit history, credit data, financial status, behavioral characteristics, social network, etc.; when nodes represent financial institutions, user credit characteristics of nodes in graph data include credit product characteristics, risk control capabilities, market performance, compliance records, etc.
[0073] In the embodiments of this specification, nodes that meet the set requirements can be used as target nodes to be evaluated. For example, nodes that originally have abnormal marks, or nodes with abnormal credit records can be directly used as target nodes.
[0074] The original graph corresponding to the graph data can be an isomorphic graph, that is, all nodes in the graph are of the same type; or it can be a heterogeneous graph, that is, the nodes contain different types.
[0075] In step 102, a plurality of subgraphs are determined in the graph data according to the target node.
[0076] By identifying and extracting a portion of nodes and edges directly or partially related to each target node from the original graph corresponding to the graph data, subgraph extraction based on the target node can be achieved, thereby determining multiple subgraphs in the graph data. In this specification, for a subgraph established based on a target node, the target node can be referred to as a central node.
[0077] In the embodiment of the present specification, the subgraph established based on each target node at least includes all directly adjacent nodes of the target node.
[0078] In step 103, the user credit feature of the first node in the subgraph is aggregated to the target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node.
[0079] Taking one of the subgraphs as an example, the subgraph is built around one of the target nodes, and the other nodes other than the target node can be called first nodes. The user credit features of each first node are aggregated to the target node through a first aggregation function, and the features of the target node are updated using the aggregated features.
[0080] Among them, the aggregation function may include general aggregation functions such as sum function, average function, maximum function, minimum function, weighted average, weighted sum, adaptive aggregation, etc.; it may also include graph metric aggregation function, which usually determines the aggregation function of node features according to graph metrics such as node degree, edge weight, path length, etc. in the graph, for example, weighted sum aggregation assigns weights to each neighbor node according to the graph metric of the node, for example, the aggregation of paths determines the weights of neighbor nodes according to the path length, etc.; in addition, it may also include aggregation functions for custom services. Those skilled in the art should understand that these aggregation functions are only examples, and any other appropriate aggregation functions may be used for feature aggregation, and the present application does not intend to limit this.
[0081] In the embodiments of the present specification, based on the characteristics of the node itself and the algorithm of the first aggregation function, the business meaning of the aggregated features aggregated to the target node can be determined. The business meaning is clear, understandable, and interpretable, rather than a high-dimensional feature with ambiguous meaning in a machine learning model such as GNN or DNN (deep neural network).
[0082] In step 104, target features are selected from the aggregated features to be input into a machine learning model, and the machine learning model is used to predict the credit risk of the user. Since in step 103, the features of the first node in each subgraph are aggregated to the target node, and the aggregated features of multiple target nodes are obtained, the graph data is converted into point table data. Therefore, it is no longer limited to processing using GNN, and other types of machine learning models, such as decision trees, support vector machines, neural networks, etc., can also be used to predict the credit risk of the user. In this step, all aggregated features can be directly input into the machine learning model as target features, or target features that meet the set requirements can be selected from the aggregated features and input into the machine learning model.
[0083] The credit risk assessment method proposed in the embodiment of this specification first selects the target node to be evaluated from multiple nodes of the graph data representing the credit characteristics of the user, and determines multiple subgraphs in the graph data according to the target node, aggregates the user credit characteristics of other nodes in the subgraph other than the target node to the target node in the subgraph to obtain aggregated characteristics, and finally selects the target characteristics input to the machine learning model from the aggregated characteristics, and the machine learning model is used to predict the credit risk of the user. On the one hand, by aggregating the user credit characteristics of other nodes in the subgraph to the target node, the aggregated characteristics with business meaning are extracted from the graph data, and the credit risk of the user is predicted based on these aggregated characteristics, which improves the interpretability of the machine learning model in the processing process; on the other hand, based on the graph data, the aggregated characteristics of each target feature are extracted for prediction, which reduces the amount of data to be processed and is also conducive to improving the interpretability of the machine learning model processing process.
[0084] In addition, since the obtained aggregated features have clear business meanings and good portability, they can be directly used as assets.
[0085] In some embodiments, a first sampling method having a preset mapping relationship with a first aggregation function may be used to sample the original graph corresponding to the graph data to obtain a subgraph corresponding to the target node.
[0086] In the embodiments of this specification, mapping relationships between multiple sampling methods and multiple aggregation functions are predefined. Among them, the sampling methods may include k-hop sampling method, Random walk sampling method, multi-hop conditional sampling method, etc. Those skilled in the art should understand that these sampling methods are only examples, and any other appropriate method may be used to sample the original graph, and this application does not intend to limit this. The above sampling methods are described below.
[0087] The k-hop sampling method starts from a node, explores the graph with a fixed width k, and collects all neighbor nodes within k hops of the node. The "hop" here refers to the edge in the graph. For example, 1-hop neighbors are directly connected nodes, 2-hop neighbors are nodes two hops away, and so on.
[0088] Taking the borrower as an example, when extracting features for assessing the borrower's credit risk, 1-hop sampling may include the borrower's immediate family members, co-borrowers, banks that the borrower frequently deals with, etc.; 2-hop sampling may be extended to the borrower's friends, friends of family members, partners of business partners, etc.; 3-hop and further sampling may be further extended to a wider network, taking into account more macro factors such as community influence and industry dynamics.
[0089] Random Walk sampling is a graph sampling method based on the random walk algorithm. In this method, starting from a node, a neighbor node is randomly selected, and then moved to the neighbor node, and this process is repeated until the preset walk length or number of steps is reached, thereby generating a random walk path. This path can be regarded as a random exploration in the graph to capture the local structure and characteristics of the nodes in the graph.
[0090] For example, if the nodes of the graph data are borrowers, we start with the borrower as the starting node, randomly select a neighbor node (such as a family member, business partner, etc.), then move to that node, and then randomly select the next neighbor node from that node, and so on. Each random walk generates a path, and the nodes on the path may include the borrower's direct and indirect contacts, as well as various relationships between them.
[0091] Multi-hop conditional sampling selectively samples information from a node’s multi-hop neighbors based on specific conditions or probabilities. This sampling strategy not only considers the node’s degree or a fixed number of hops, but may also consider the node’s attributes, edge weights, graph structural features, and other contextual information related to the current task.
[0092] Still taking the borrower as an example, sampling conditions can be set according to the needs of credit risk assessment, including the credit score of neighbor nodes, transaction amount with the borrower, borrower's social network influence, etc. Starting from the borrower as the starting node, according to the set sampling conditions, nodes are selectively sampled from its direct neighbors (such as family members, business partners, etc.), and then according to the same conditions, continue to sample from the neighbors of these neighbors, and perform a multi-hop sampling process. After that, the borrower and all sampled multi-hop neighbor nodes and the edges between them are extracted to form a subgraph.
[0093] It can be seen from the above that the subgraphs corresponding to the target nodes reached by different sampling methods are different.
[0094] In the embodiments of this specification, the aggregation function may include general aggregation functions such as sum function, average function, maximum function, minimum function, weighted average, weighted sum, adaptive aggregation, etc.; it may also include graph metric aggregation function, which usually determines the aggregation function of node features according to graph metrics such as node degree, edge weight, path length, etc. in the graph, for example, weighted sum aggregation assigns weights to each neighbor node according to the graph metric of the node, for example, the aggregation of paths determines the weights of neighbor nodes according to the path length, etc.; in addition, it may also include aggregation functions for custom services. Those skilled in the art should understand that these aggregation functions are only examples, and any other appropriate aggregation functions may be used for feature aggregation, and this application does not intend to limit this.
[0095] By establishing a mapping relationship between the sampling method and the aggregation function, the combination of different aggregation functions and sampling methods will affect the representation of the aggregated features, thereby affecting downstream tasks (such as credit risk prediction based on aggregated features).
[0096] For example, if the average function has a mapping relationship with k-hop sampling, that is, the average function and k-hop sampling form a combination, the resulting aggregated features focus on information in the local area and will reflect the collective attributes of the direct neighbors of the central node. It is suitable for situations where the feature importance of neighboring nodes is relatively equal or the structure of the graph is relatively regular.
[0097] For another example, the weighted sum function can be set to have a mapping relationship with Random Wa lk sampling, that is, the weighted sum function and Random Wa lk sampling form a combination, which can capture the long-term dependencies in the graph and reflect the importance of different neighbors through the weights of the edges. The aggregated features will contain information about the nodes encountered during the random walk, which may be far away. It is suitable for situations where the structure of the graph is more complex and distant neighbors may also have a significant impact on the characteristics of the nodes.
[0098] For another example, a mapping relationship can be established between the graph metric (such as node degree) aggregation function and multi-hop conditional sampling to form a combination. Specifically, the borrower's direct and indirect contacts, such as family members, business partners, etc., can be selected based on the borrower's credit and transaction amount. When aggregating the characteristics of neighbor nodes, it can be assumed that nodes with higher node degrees have a greater impact on the central node, so the node can be given a greater weight. Then, the characteristics of the neighbor nodes are combined by weighted summation, thereby obtaining the characteristics of the borrower's social and economic environment.
[0099] It can be seen that by combining different sampling methods and different aggregation functions, aggregation features with different business meanings can be obtained.
[0100] Figure 2 This is a flow chart of a method for determining a mapping relationship between a sampling method and an aggregation function provided by an exemplary embodiment. The method includes steps 201 to 203.
[0101] In step 201, each sampling method among multiple sampling methods is combined with each aggregation function among multiple aggregation functions one by one to obtain multiple groups of sampling aggregation pairs.
[0102] like Figure 3 As shown, Figure 3 The left side shows a variety of sampling methods, and the right side shows a variety of aggregation functions. By combining each sampling method with the first aggregation function one by one, multiple groups of sampling aggregation pairs can be obtained.
[0103] For example, the obtained sampling aggregation pairs may include k-hop sampling-sum function, Random wal k-average function, multi-hop conditional sampling-node degree aggregation function, etc.
[0104] In step 202, the sample graph data is processed using each group of sampling aggregation pairs to obtain sample features.
[0105] The sample graph data is graph data in which the true values of the credit risk assessment results are marked on the sample nodes.
[0106] Each group of sampling and aggregation pairs is used to sample the sample graph data to obtain a subgraph, and then the features are aggregated to obtain the sample aggregation features of each sample node.
[0107] In step 203, a mapping relationship between a sampling method and an aggregation function is determined based on the credit risk prediction result corresponding to the sample aggregation feature and the corresponding true value.
[0108] By inputting the sample aggregation features of each sample into the machine learning model used to predict the credit risk of the user, the credit risk prediction results of each sample can be obtained. According to the true value of the credit risk prediction results marked by each sample, it can be determined which sampling aggregation pair can be applied to obtain a better credit risk assessment effect. Therefore, a mapping relationship can be established between the sampling methods and aggregation functions in these sampling aggregation pairs that meet the requirements. For example, the requirement that is met can be that the difference between the prediction result and the true value is less than a set threshold.
[0109] In the disclosed embodiment, by determining sampling aggregation pairs whose credit risk assessment effect meets the requirements and establishing a mapping relationship between the sampling methods and aggregation functions in these sampling aggregation pairs, it is ensured that the obtained aggregation features have a good credit risk prediction effect.
[0110] In some embodiments, the user credit feature of a node in the graph data includes multiple sub-features. Taking the node representing a borrower as an example, the sub-features included in the user credit feature of the node include credit history, credit data, financial status, behavioral characteristics, social network, etc.
[0111] Each sub-feature can be aggregated to the target node to obtain a sub-aggregation feature, and the aggregate feature of the target node can be obtained based on each sub-aggregation feature.
[0112] Since each sub-feature may contain different aspects or patterns of the original feature, aggregating these sub-features separately can help the machine learning model capture more detailed information and make the model's decision-making process more transparent, further improving the interpretability of the model processing and helping to discover potential risk points, thereby improving the assessment of credit risk.
[0113] In some embodiments, the aggregate feature may be subjected to abnormal sample detection, and the abnormal samples may be eliminated, and then the aggregate feature may be updated using the elimination result.
[0114] In some examples, detection methods such as Isolation Forest, Local Outlier Factor, and DBSCAN (Density-Based Spatial Classification of Applications with Noise) can be used to detect abnormal samples.
[0115] By removing abnormal samples, the impact of outliers on feature importance can be reduced and the amount of calculation can be reduced.
[0116] In some embodiments, the aggregated features are screened to obtain target features whose correlation with the indicator to be evaluated meets the set requirements.
[0117] In some examples, recursive feature elimination (RFE), model-based feature selection, minimum redundancy maximum relevance (mRMR), LASSO (Least Absolute Shr inkage and Selection Operator), principal component analysis, Boruta and other methods can be used for feature screening to select target features with a higher correlation with user credit risk from the aggregated features.
[0118] Since the aggregate feature contains multiple sub-aggregate features, the correlation between different sub-aggregate features and the evaluation indicator of user credit risk is different. By selecting sub-aggregate features with high correlation with the evaluation indicator and inputting the target features composed of these sub-aggregate features into the machine learning model for predicting user credit risk, the low-correlation features are excluded, the impact of noise is reduced, and it helps the model focus on important information, thereby improving the accuracy of the prediction and reducing the risk of overfitting.
[0119] In some embodiments, the graph data may be cleaned before selecting the target node, and the data cleaning includes at least one of excluding strong linear correlation features and removing hot spot data.
[0120] Since features in graph data may have multicollinearity problems, that is, there is a high linear relationship between two or more features, such graph data may lead to unstable model training, increase computational complexity, and even affect the generalization ability of the model. By excluding strongly linearly correlated features, such as by calculating the correlation coefficient between features (such as the Pearson correlation coefficient) to identify strongly linearly correlated features and excluding these highly correlated features, it helps to improve the stability and generalization ability of the model, and at the same time improve the prediction accuracy of the model.
[0121] Hot data refers to data points that appear too frequently in graph data. They may be caused by errors in data collection or processing. Hot data may mislead the analysis of the graph and affect the performance of the model. By analyzing the degree of nodes, the weight of edges, or other indicators to identify hot data, and then removing these data points according to certain criteria (such as degree thresholds), the noise in the data set can be reduced and the representativeness of the data set can be improved, which may improve the accuracy and reliability of subsequent analysis or models.
[0122] Figure 4 is a flowchart of another credit risk assessment method provided by an exemplary embodiment. Figure 4 As shown, the method includes steps 401-405.
[0123] In step 401, data cleaning is performed on the graph data.
[0124] In this embodiment, the graph data is graph data applied to the credit field. The nodes in the graph data can represent entities such as borrowers and financial institutions, and the edges represent credit relationships, transaction relationships, etc. between entities. When the nodes represent borrowers, the characteristics of the nodes in the graph data can include user credit characteristics, including the user's credit history, credit data, financial status, behavioral characteristics, social network, etc.; when the nodes represent financial institutions, the user credit characteristics of the nodes in the graph data include credit product characteristics, risk control capabilities, market performance, compliance records, etc.
[0125] In the graph data, nodes that meet the set requirements can be used as target nodes to be evaluated. For example, nodes that originally have abnormal marks, or nodes with abnormal credit records can be directly used as target nodes.
[0126] In step 402, the original graph corresponding to the graph data is sampled using a first sampling method to obtain a subgraph corresponding to the target node.
[0127] The first sampling method can be any one of a k-hop sampling method, a random water sampling method, and a multi-hop conditional sampling method.
[0128] By sampling each target node, a subgraph corresponding to each target node can be established. The target node can also be called the central node of the subgraph.
[0129] In step 403, each sub-feature of the first node other than the target node in the subgraph is aggregated to the target node by a first aggregation function, and an aggregate feature of the target node is obtained according to each aggregated sub-aggregation feature.
[0130] The first aggregation function and the first sampling method have a preset mapping relationship, that is, after determining which method is used in the first sampling method, it is possible to determine which aggregation function is used.
[0131] For example, in the disclosed example, a mapping relationship is established in advance between k-hop sampling and an average function, a mapping relationship is established between Random Wa lk and a weighted sum function, and a mapping relationship is established between multi-hop conditional sampling and a node degree aggregation function. In this case, if the Random Wa lk sampling method is used for sampling in step 402, the weighted sum function is used to aggregate the features of the first node in step 403 to obtain the aggregated features of the target node.
[0132] In step 404, abnormal sample detection is performed on the aggregated features, and abnormal samples are removed.
[0133] In step 405, the aggregated features are screened to obtain target features whose correlation with the indicator to be evaluated meets the set requirements.
[0134] In this embodiment of the specification, by aggregating user credit features of other nodes in the subgraph to the target node, aggregate features with business significance are extracted from the graph data, and the user's credit risk is predicted based on these aggregate features, thereby improving the interpretability of the machine learning model during the processing process, and by aggregating user credit features of other nodes in the subgraph to the target node, thereby extracting aggregate features with business significance from the graph data, and predicting the user's credit risk based on these aggregate features, thereby improving the interpretability of the machine learning model during the processing process.
[0135] Moreover, by determining the sampling aggregation pairs whose credit risk assessment effects meet the requirements and establishing a mapping relationship between the sampling methods and aggregation functions in these sampling aggregation pairs, it is ensured that the obtained aggregation features have a better credit risk prediction effect.
[0136] In addition, when the scale of graph data is huge, the method proposed in this embodiment of the specification can efficiently extract highly relevant and explainable features directly from the graph data, reducing the difficulty of feature engineering.
[0137] Figure 5 FIG. 1 is a block diagram of a credit risk assessment device provided by an exemplary embodiment. Figure 5 As shown, the device comprises:
[0138] A selection unit 501 is used to select a target node to be evaluated from multiple nodes in the graph data, wherein the characteristics of the nodes in the graph data include user credit characteristics;
[0139] A determining unit 502, configured to determine a plurality of subgraphs in the graph data according to the target node;
[0140] An aggregation unit 503, configured to aggregate the user credit feature of the first node in the subgraph to the target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node;
[0141] The prediction unit 504 is used to select target features from the aggregated features to be input into a machine learning model, where the machine learning model is used to predict the credit risk of the user.
[0142] In some embodiments, the determining unit is specifically configured to:
[0143] Determine the subgraph corresponding to the target node by adopting a first sampling method;
[0144] Among them, the first sampling method and the first aggregation function have a preset mapping relationship.
[0145] In some embodiments, the apparatus further comprises a mapping unit configured to:
[0146] Combining each sampling method of the plurality of sampling methods with each aggregation function of the plurality of aggregation functions one by one to obtain a plurality of groups of sampling aggregation pairs;
[0147] Using each group of sampling aggregation pairs to process the sample graph data respectively, to obtain sample aggregation features, wherein the sample graph data is graph data with true values of credit risk assessment results marked on sample nodes;
[0148] According to the credit risk prediction results corresponding to the sample aggregation characteristics and the corresponding true values, the mapping relationship between the sampling method and the aggregation function is determined.
[0149] In some embodiments, the user credit feature of the node includes a plurality of sub-features;
[0150] The polymeric unit is specifically used for:
[0151] Aggregating each sub-feature of the first node to the target node to obtain sub-aggregated features;
[0152] The aggregate feature is obtained according to each sub-aggregate feature.
[0153] In some embodiments, the device further comprises an abnormality elimination unit, configured to:
[0154] Performing abnormal sample detection on the aggregated features and removing abnormal samples;
[0155] The aggregated features are updated using the eliminated results.
[0156] In some embodiments, the device further comprises a screening unit for:
[0157] The aggregated features are screened to obtain target features whose correlation with the indicator to be evaluated meets the set requirements.
[0158] In some embodiments, the device further comprises a data cleaning unit, configured to:
[0159] The graph data is cleaned, wherein the data cleaning includes at least one of excluding strong linear correlation features and removing hot spot data.
[0160] Figure 6 is a schematic structural diagram of a device provided by an exemplary embodiment. Figure 6At the hardware level, the device includes a processor 602, an internal bus 604, a network interface 606, a memory 608, and a non-volatile memory 610, and may also include hardware required for other services. One or more embodiments of this specification may be implemented based on software, such as the processor 602 reading the corresponding computer program from the non-volatile memory 610 into the memory 608 and then running it. Of course, in addition to the software implementation, one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.
[0161] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which may be in the form of a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver, a game console, a tablet computer, a wearable device or a combination of any of these devices.
[0162] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0163] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0164] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.
[0165] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0166] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0167] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0168] It should be understood that although the terms first, second, third, etc. may be used to describe various information in one or more embodiments of this specification, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0169] The above description is merely a preferred embodiment of one or more embodiments of the present specification and is not intended to limit one or more embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present specification shall be included in the scope of protection of one or more embodiments of the present specification.
Claims
1. A credit risk assessment method, comprising: Selecting a target node to be evaluated from multiple nodes in the graph data, wherein the characteristics of the nodes in the graph data include user credit characteristics; Determine a plurality of subgraphs in the graph data according to the target node; aggregating the user credit feature of the first node in the subgraph to the target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node; A target feature is selected from the aggregated features and input into a machine learning model, where the machine learning model is used to predict the credit risk of the user.
2. According to the method of claim 1, determining a plurality of subgraphs in the graph data according to the target node comprises: Determine the subgraph corresponding to the target node by adopting a first sampling method; Among them, the first sampling method and the first aggregation function have a preset mapping relationship.
3. The method according to claim 2, further comprising: Combining each sampling method of the plurality of sampling methods with each aggregation function of the plurality of aggregation functions one by one to obtain a plurality of groups of sampling aggregation pairs; Using each group of sampling aggregation pairs to process the sample graph data respectively, to obtain sample aggregation features, wherein the sample graph data is graph data with true values of credit risk assessment results marked on sample nodes; According to the credit risk prediction results corresponding to the sample aggregation characteristics and the corresponding true values, the mapping relationship between the sampling method and the aggregation function is determined.
4. The method according to any one of claims 1 to 3, wherein the user credit feature of the node comprises a plurality of sub-features; Aggregating the user credit feature of the first node in the subgraph to the target node in the subgraph to obtain the aggregated feature includes: Aggregating each sub-feature of the first node to the target node to obtain sub-aggregated features; The aggregate feature is obtained according to each sub-aggregate feature.
5. The method according to claim 4, further comprising: Performing abnormal sample detection on the aggregated features and removing abnormal samples; The aggregated features are updated using the eliminated results.
6. The method according to claim 5, further comprising: The aggregated features are screened to obtain target features whose correlation with the indicator to be evaluated meets the set requirements.
7. The method according to claim 1, further comprising: The graph data is cleaned, wherein the data cleaning includes at least one of excluding strong linear correlation features and removing hot spot data.
8. A credit risk assessment device, comprising: A selection unit, configured to select a target node to be evaluated from a plurality of nodes in the graph data, wherein the characteristics of the nodes in the graph data include user credit characteristics; A determination unit, configured to determine a plurality of subgraphs in the graph data according to the target node; an aggregation unit, configured to aggregate the user credit feature of a first node in the subgraph to a target node in the subgraph through a first aggregation function to obtain an aggregate feature, wherein the first node is a node in the subgraph other than the target node; A prediction unit is used to select target features from the aggregated features to be input into a machine learning model, where the machine learning model is used to predict the credit risk of a user.
9. An electronic device, comprising: processor; a memory for storing processor-executable instructions; The processor implements the method according to any one of claims 1 to 7 by running the executable instructions.
10. A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.
11. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.
