Credit card behavior feature recognition method, device and equipment and storage medium

Abstract

The invention discloses a credit card behavior feature recognition method, device and equipment and a storage medium, and relates to the technical field of finance, and the method comprises the steps: obtaining transaction flow data, carrying out the preprocessing of the transaction flow data, obtaining a plurality of data samples, and enabling the data samples to comprise labeled samples and non-labeled samples, the marked sample has a credit card behavior characteristic type mark; performing node state calculation on each marked sample to generate a node state sequence corresponding to each marked sample; converting the node state sequence into a multi-view image, and training a multi-view image representation learning model based on the multi-view image; performing credit card behavior feature recognition on unmarked samples through the trained multi-view image representation learning model; the credit card use behavior characteristics can be accurately identified in a short time.

Description

Technical Field

[0001] The present application relates to the field of financial technology, and in particular to a credit card behavior feature recognition method, device, equipment and storage medium. Background Art

[0002] In the financial field, mining bank credit card transaction flow data, understanding and identifying customers' behavioral characteristics in using credit cards, has important practical significance for risk management, providing personalized services and marketing strategies, and maximizing bank profits.

[0003] At present, the research methods for understanding and identifying the characteristics of customers' credit card behavior, on the one hand, need to rely on the customer's years of work, marital status, income and other private data for analysis. On the other hand, feature engineering methods that rely on domain expertise and expert knowledge such as credit limit utilization rate and repayment rate, such as logistic regression and support vector machines, are used; the expressiveness and descriptive ability of extracted features are relatively limited. In addition, many studies focus on identifying credit card behavior characteristics for long-term or entire customers, which is inconsistent with the fact that customer credit card behavior characteristics are complex and change over time in a fine-grained manner.

[0004] CN114722935A provides a credit card customer data segmentation method and device. The method first pre-processes the original credit card customer data according to the segmentation requirements to obtain a target credit card customer data set. Then, the target credit card customer data set is used to create a credit card customer data KD-tree; finally, the target credit card customer data set is segmented using a spatial search algorithm that combines the credit card customer data KD-tree with the DBSCAN algorithm. On the one hand, the patent calculates the data variance on different dimensions based on customer attributes such as customer income and consumption amount credit card, and selects the largest variance value as the segmentation dimension. This ability to extract relevant features based on indicators defined by domain expertise is relatively limited. On the other hand, the patent is segmented at the granularity of the entire credit card customer, which cannot reflect the customer's credit card usage behavior characteristics in the short term, and thus it is difficult to make timely and accurate risk management and provide personalized services.

[0005] CN116385145A proposes a method for predicting credit card customer behavior. First, in terms of feature engineering, based on historical customer credit card overdue records, data analysis and processing techniques such as missing value filling, unique hot encoding, chi-square analysis, correlation analysis, and data binning are adopted to generate a modeling data table. Then, an XGBoost classification model based on clustering and bagging is constructed, and the training set is used for training. Finally, the pre-trained credit card customer behavior prediction model is obtained by adjusting parameters. On the one hand, this patent studies the feature representation learning of positive and negative samples of customer complaint behavior, which is a binary classification problem and is not suitable for the problem of identifying complex credit card usage behavior features of customers in the short term. On the other hand, the invention uses traditional machine learning methods, which are relatively limited in the ability to extract advanced features compared to artificial intelligence algorithms such as deep learning and representation learning, and thus it is difficult to effectively solve complex multi-classification problems.

[0006] In view of this, how to solve the problem of identifying the characteristics of credit card usage behavior in the short term has become an urgent problem to be solved. Summary of the invention

[0007] In view of this, the main purpose of this application is to provide a credit card behavior feature recognition method, device, equipment and storage medium to accurately identify credit card usage behavior characteristics in a short period of time.

[0008] The first aspect of the present application provides a credit card behavior feature recognition method, the method comprising:

[0009] Acquire transaction flow data, and pre-process the transaction flow data to obtain multiple data samples, the data samples include labeled samples and unlabeled samples, and the labeled samples are labeled with credit card behavior feature types;

[0010] By calculating the node state of each labeled sample, a node state sequence corresponding to each labeled sample is generated;

[0011] Convert the node state sequence into a multi-view graph, and train a multi-view graph representation learning model based on the multi-view graph;

[0012] Credit card behavior features are recognized for unlabeled samples through the trained multi-view graph representation learning model.

[0013] In some implementations of the first aspect of the present application, the transaction flow data is preprocessed to obtain a plurality of data samples, including:

[0014] Divide the transaction flow data into multiple continuous and equal-length data segments to obtain multiple data samples;

[0015] A plurality of data samples for model training are randomly selected from a plurality of data samples, and credit card behavior feature type tags are added to the plurality of data samples for model training to obtain a plurality of labeled samples.

[0016] In some implementations of the first aspect of the present application, the transaction flow data is composed of a series of transaction points sorted by time; the transaction flow data is divided into a plurality of continuous and equal-length data segments, including:

[0017] Starting from the first transaction point of the transaction flow data, through a time window of preset length, continuously slide to the last transaction point of the transaction flow data with a step of preset length to obtain multiple continuous and equal-length data fragments, and the data fragments are the same length as the time window.

[0018] In some implementations of the first aspect of the present application, a node state sequence corresponding to each marked sample is generated by performing node state calculation on each marked sample, including:

[0019] For each marked sample, the node status of each transaction point in the marked sample is described by the credit limit and the remaining limit;

[0020] The credit limit range is divided into m credit limit discretization intervals, and the remaining limit range is divided into n remaining limit discretization intervals, and the m credit limit discretization intervals and the n remaining limit discretization intervals constitute m*n discretization node states, where m and n are both positive integers;

[0021] Based on the node state of each transaction point, each transaction point is mapped to the corresponding discretized node state to obtain the node state sequence corresponding to the labeled sample.

[0022] In some implementations of the first aspect of the present application, the multi-view graph includes: a frequency graph and a time graph, and converting the node state sequence into the multi-view graph includes:

[0023] Based on the state transition frequency between different discretized node states, the node state sequence is converted into a frequency graph, wherein the nodes of the frequency graph represent the discretized node states, and the edges of the frequency graph represent the state transition frequency between different discretized node states;

[0024] Based on the average state transition time between different discretized node states, the node state sequence is converted into a time graph, where the nodes of the time graph represent the discretized node states, and the edges of the time graph represent the average state transition time between different discretized node states.

[0025] In some implementations of the first aspect of the present application, the multi-view graph representation learning model includes: a graph convolutional network and a classifier;

[0026] The graph convolutional network is used to extract and pool the node features of each node in the multi-view graph, flatten the node features of each node into one dimension and map them to the sample label space to obtain the probability distribution of data samples on different credit card customer behavior feature categories;

[0027] The classifier assigns credit card customer behavior feature category labels to data samples based on the feature representation extracted by the graph convolutional network.

[0028] In some implementations of the first aspect of the present application, the graph convolutional network includes: a graph convolution layer, a graph embedding pooling layer, a flattening layer, and a fully connected layer;

[0029] The graph convolution layer is used to extract the features of each node in the multi-view graph to obtain the initial feature representation of each node in the multi-view graph;

[0030] The graph embedding pooling layer pools the initial feature representation of each node through weighted convex combination to obtain a new feature representation of each node in the multi-view graph;

[0031] The flattening layer is used to flatten the new feature representation of each node after pooling into one dimension;

[0032] The fully connected layer is used to map the new feature representation of each flattened node to the sample label space to obtain the probability distribution of the data sample on different credit card customer behavior feature categories.

[0033] The second aspect of the present application provides a credit card behavior feature recognition device, the device comprising:

[0034] A data processing module is used to obtain transaction flow data and pre-process the transaction flow data to obtain multiple data samples, wherein the data samples include marked samples and unmarked samples, and the marked samples are marked with credit card behavior feature type;

[0035] A sequence generation module, used to generate a node state sequence corresponding to each labeled sample by performing node state calculation on each labeled sample;

[0036] A model training module, used to convert the node state sequence into a multi-view graph, and train a multi-view graph representation learning model based on the multi-view graph;

[0037] The feature recognition module is used to identify credit card behavior features of unlabeled samples through the trained multi-view graph representation learning model.

[0038] The third aspect of the present application provides a credit card behavior feature recognition device, which includes a memory and a processor, and the processor is used to execute the program stored in the memory and run any credit card behavior feature recognition method provided in the first aspect of the present application.

[0039] The fourth aspect of the present application provides a readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, each step of the credit card behavior feature recognition method as provided in any one of the first aspects of the present application is implemented.

[0040] The technical solution provided by this application has the following beneficial effects:

[0041] In the technical solution provided in the present application, first, starting from the original transaction flow data, the transaction flow data is preprocessed, and the preprocessed data samples include labeled samples with labels and unlabeled unlabeled samples. The credit card behavior feature type labels in the labeled samples can provide supervision information to build an effective model; then, the node state is calculated based on each transaction point of the labeled samples to obtain a node state sequence, so that the continuous transaction behavior is divided into a state transition process, thereby accurately reflecting the customer's credit card usage behavior characteristics from the perspective of short-term changes; thereafter, the node state sequence is converted into a multi-view graph, and the multi-view graph representation learning model is trained through the multi-view graph, which can extract the characteristics of credit card usage behavior from different perspectives, so that the model learns more comprehensive behavioral characteristics instead of relying on a single dimension, so that it has a strong fine-grained analysis capability in time changes, which is suitable for short-term behavior recognition; finally, the unlabeled samples are feature recognized by the trained model, and the high expression ability of the multi-view graph representation learning model is used for inference to adapt to the rapidly changing behavioral characteristics in the short term. It can be seen that this application can accurately identify the characteristics of credit card transaction behavior in a short period of time through data preprocessing, node state calculation, node state sequence and multi-view graph construction, combined with a multi-view graph representation learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart of a credit card behavior feature recognition method provided in an embodiment of the present application;

[0043] Figure 2 A flowchart of another credit card behavior feature recognition method provided in an embodiment of the present application;

[0044] Figure 3 A flowchart of another credit card behavior feature recognition method provided in an embodiment of the present application;

[0045] Figure 4 A flowchart of another credit card behavior feature recognition method provided in an embodiment of the present application;

[0046] Figure 5 A framework diagram of a fine-grained credit card customer behavior feature recognition system based on multi-view graph representation learning provided in an embodiment of the present application;

[0047] Figure 6 A schematic diagram of the structure of a credit card behavior feature recognition device provided in an embodiment of the present application;

[0048] Figure 7 A structural diagram of a credit card behavior feature recognition device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] Glossary:

[0051] Fine-grained credit card behavior characteristics: Credit card behavior characteristics are a complex combination of customers' consumption and repayment behaviors and habits in the process of using credit cards. In the embodiments of the present application, short-term credit card customer behavior characteristics are defined, such as half a year. Credit card customer behavior characteristics are divided into four categories of characteristics from weak to strong according to the ability of consumption contribution and repayment risk, namely high contribution and low risk, low contribution and low risk, high contribution and high risk, and low contribution and high risk. It is worth noting that the overall credit card customer performance is a growth vector composed of these four types of behavioral characteristics and evolves over time. For example, if a customer has been using a credit card with high contribution and low risk characteristics for a long time, that is, most of the vectors are high contribution and low risk characteristics, then the customer can be called a high-end customer. Similarly, potential customers, mid-range customers and low-end customers correspond to the other three characteristics respectively. Therefore, the fine-grained credit card behavior feature recognition method proposed in the present invention can be expanded to solve the problem of credit card customer classification.

[0052] Multi-view graph representation learning: Nowadays, many real-world problems can be described and quantified based on effective graph data. However, most existing graph representation learning methods only focus on processing one graph. However, real-world graph data is much more complex. Not only does each node have its own feature information, but different perspectives often contain different types of structural information. If features are only characterized from each isolated perspective, incorrect or even contradictory recognition results may occur. To this end, it is necessary to fuse multiple views into a complete and systematic whole, extract information between multiple views, and model the relationship between multiple views. Multi-view graph representation learning is an effective method for deep learning and feature extraction of complex multi-view graph data.

[0053] The embodiments of the present application provide a credit card behavior feature recognition method, device, equipment and storage medium.

[0054] See also Figure 1 As shown, the embodiment of the present application provides a credit card feature recognition method, which may specifically include the following steps:

[0055] S101: Acquire transaction flow data, and pre-process the transaction flow data to obtain multiple data samples, where the data samples include marked samples and unmarked samples, and the marked samples are marked with credit card behavior feature type.

[0056] In the embodiments of the present application, transaction flow data refers to records containing credit card transactions, typically including transaction time, amount, credit limit, remaining balance and other information. The transaction flow data can reflect the specific behavior of customers using credit cards, and a data sample refers to a data fragment in the transaction flow data.

[0057] Among them, data samples are divided into labeled samples and unlabeled samples. Labeled samples are data samples marked with credit card behavior feature types and used for training multi-view graph representation learning models. Unlabeled samples are data samples not marked with credit behavior feature types and used for model classification prediction.

[0058] See also Figure 2 As shown, in some implementations of the embodiments of the present application, the process of preprocessing the transaction flow data to obtain multiple data samples may specifically include the following steps:

[0059] S201: Divide the transaction flow data into a plurality of continuous and equal-length data segments to obtain a plurality of data samples.

[0060] In an embodiment of the present application, under the premise of protecting the security of credit card customers and privacy information such as name, ID number and income, in order to identify the behavioral characteristics of customers in using credit cards in the short term, the customer's credit card transaction flow data is divided into a series of continuous and equal-length data segments based on data segmentation, and each data segment is regarded as a data sample.

[0061] The data segmentation method can be based on a given credit card transaction flow data T with a length of l. <p1,p2,...,p l >, in the transaction flow data T, each element represents a transaction record containing timestamp ti, credit limit ci and remaining limit ri, that is, pi = (ti, ci, ri), and a series of data fragments of transaction flow data, that is, data samples τ, are obtained. Therefore, the goal of data segmentation can be formalized as T = <id,τ1,...,τ n >, where id represents the identifier of the customer.

[0062] S202: Randomly select multiple data samples for model training from multiple data samples, and add credit card behavior feature type tags to the multiple data samples for model training to obtain multiple labeled samples.

[0063] In an embodiment of the present application, multiple data samples obtained through data segmentation processing constitute a sample set. Then, multiple data samples for model training are selected from the sample set by randomly selecting samples. At the same time, they are manually marked as corresponding credit card behavior feature types based on expert knowledge to guide the feature learning and optimization of the model.

[0064] In some implementations of the embodiments of the present application, the transaction flow data is composed of a series of transaction points sorted by time. Specifically, the transaction flow data is a plurality of records arranged in chronological order, and each record represents a transaction point. Each transaction point records a transaction event that occurs at a specific time, and these transaction events are arranged together in chronological order to form a time series. On this basis, the transaction flow data is divided into a plurality of continuous and equal-length data segments in step S201, which can be specifically implemented in the following manner:

[0065] Starting from the first transaction point of the transaction flow data, through a time window of preset length, continuously slide to the last transaction point of the transaction flow data with a step of preset length to obtain multiple continuous and equal-length data fragments, and the data fragments are the same length as the time window.

[0066] Specifically, first, given a credit transaction flow data T, a time window of length 1s is defined, which refers to a time range of fixed length, such as 5 minutes, 1 hour, 1 day, etc. The length of the time window determines the time span of each data fragment. Then, it starts from the first transaction point of the credit transaction flow data T, that is, the earliest transaction recorded in the transaction flow data T, and slides continuously to the last point with a step of length 1s, thereby obtaining multiple transaction flow data fragments. The time span of each data fragment is the same as the length of the preset time window, which is 1s. It can be understood that the transaction flow data of other customers are also pre-processed based on the method provided in step S201 to obtain different data samples.

[0067] S102: Generate a node state sequence corresponding to each marked sample by performing node state calculation on each marked sample.

[0068] In an embodiment of the present application, the node status is used to describe the account status of a credit card at a specific point in time, such as the credit limit and the remaining balance, and the node status sequence is a sequence of all node statuses of a customer within a period of time arranged in chronological order.

[0069] By generating node state sequences, fine-grained behavioral information can be provided for each labeled sample, such as the credit limit, the high and low change trend of the remaining balance, etc., which helps the model capture subtle behavioral differences in the time dimension, thereby improving the recognition accuracy of subsequent credit card behavior characteristics.

[0070] See also Figure 3 As shown, in some implementations of the embodiments of the present application, in S102, node state calculation is performed on each marked sample to generate a node state sequence corresponding to each marked sample, which may specifically include the following steps:

[0071] S301: For each marked sample, describe the node status of each transaction point in the marked sample through the credit limit and the remaining limit.

[0072] In the embodiment of the present application, based on the preprocessing of the transaction flow data, each element in the data sample τ can be represented by a two-tuple pi=<ci,ri> To represent the node state corresponding to the transaction point pi, ci and ri represent the credit limit and remaining limit of the current transaction node respectively. The credit limit refers to the maximum loan amount that the bank can give to the user, and the remaining limit refers to the user's available limit at a certain moment, that is, the remaining amount after deducting the used part from the credit limit.

[0073] S302: Divide the credit limit range into m discrete intervals of the credit limit, divide the remaining limit range into n discrete intervals of the remaining limit, and form m*n discrete node states by the m discrete intervals of the credit limit and the n discrete intervals of the remaining limit, where m and n are both positive integers.

[0074] In the embodiment of the present application, the two attributes of the node state are continuous, which results in many different node states even for short transaction flow data, thereby increasing the complexity of the model and hindering the analysis and understanding of fine-grained credit card customer behavior. Therefore, the present invention adopts an equal frequency discretization method to limit the number of node states, that is, the attribute values ​​of the two dimensional data of credit limit and remaining limit are sorted respectively, and then the data is discretized according to a given number of intervals to ensure that the number of samples in each interval is basically the same, so that the data becomes evenly distributed.

[0075] If the number of discretized intervals of the credit limit and the remaining limit are m and n respectively, then the total number of node states is N = m × n. In this case, the degree of discretization of the node state, that is, the value of m and n, will have an important impact on the subsequent multi-view construction process and the credit card customer behavior feature recognition model.

[0076] Assume that the credit limit range is 0 to 10,000 yuan, divided into 5 intervals, that is, m = 5; the remaining limit range is 0 to 10,000 yuan, divided into 4 intervals, that is, n = 4. Given that equal frequency discretization is based on the distribution of data, the data range is divided into intervals with approximately the same number of samples in each interval. Therefore, the situation after division is as follows: First, for the equal frequency division of the credit limit interval, after statistics of the data in the credit limit range of 0-10,000 yuan, assuming that the data points are concentrated in the low credit limit part, with the goal of containing approximately equal numbers of data points, the five credit limit intervals after division may be: 0-1500 interval, 1500-3000 interval, 3000-4500 interval, 4500-7000 interval and 7000-10000 interval, where the amount of data in each interval is close to equal, but the interval width is not necessarily equal. Secondly, for the equal frequency division of the remaining credit interval, assuming that the remaining credit data is evenly distributed, the four intervals after the division may also be the same as the result of the equal width discretization division. For example, the four remaining credit intervals after the division may be: 0-2500 interval, 2500-5000 interval, 5000-7500 interval and 7500-10000 interval. Finally, the credit limit interval and the remaining credit interval are combined to form 5*4 discretized node states. Each discretized node state represents a combination of a credit limit interval and a remaining credit interval. For example: the credit limit is 0-1500 yuan, the remaining credit limit is 0-2500 yuan, which represents a discretization state; the credit limit is 4500-7000 yuan, and the remaining credit limit is 5000-7500 yuan, which represents another discretized node state. It can be seen that the present application limits the number of node states by the frequency discretization method, which can avoid the increase of model complexity.

[0077] S303: Based on the node state of each transaction point, each transaction point is mapped to a corresponding discretized node state to obtain a node state sequence corresponding to the marked sample.

[0078] In the embodiment of the present application, a transaction point can be understood as a transaction record, each transaction record represents a specific transaction of a user at a certain moment, and the node state of the transaction record is composed of its credit limit and remaining limit at that moment. For example, the credit limit corresponding to a certain transaction point is 5,000 yuan, and the remaining limit is 3,000 yuan. In S302, the credit limit and the remaining limit are divided into several discrete interval combinations, and each combination represents a discretized node state. For example, after the transaction records in the credit card transaction flow data set are discretized with equal frequency, the combination of the credit limit in the range of 2,000-4,000 yuan and the remaining limit in the range of 5,000-7,500 yuan can be defined as a discretized node state, thereby forming a series of predefined node states. The node state of each transaction point, that is, its actual credit limit and remaining limit are matched with these predefined discretized node states, and the corresponding discretized node state is determined according to the actual credit limit and remaining limit of the transaction point. For example, if the credit limit of a transaction point is 2,500 yuan and the remaining balance is 6,000 yuan, it will be mapped to the discretized node state of "credit limit between 2,000 and 4,000 yuan and remaining balance between 5,000 and 7,500 yuan". Finally, after mapping all transaction points in each labeled sample, a set of discretized node states can be obtained, which are arranged in chronological order to form a node state sequence corresponding to the labeled sample.

[0079] S103: Convert the node state sequence into a multi-view graph, and train a multi-view graph representation learning model based on the multi-view graph.

[0080] In an embodiment of the present application, the node state sequence is converted into a multi-perspective graph, and the customer's transaction behavior dynamics are modeled into a graph data structure from multiple perspectives, so that the relationships under different perspectives can be represented by a graph structure, and then can be learned and understood by a multi-perspective graph representation learning model.

[0081] Among them, the multi-view graph is a graph structure containing multiple perspectives or attributes, which is used to express the relationship between the same object in different feature dimensions. In the embodiment of the present application, it is specifically used to describe the different features and relationships of the node status. The multi-view graph representation learning model is a machine learning model trained based on the above multi-view graph, which is used to learn and extract high-dimensional features from the multi-view graph to obtain low-dimensional and expressive feature representations. By modeling and exploring the relationships between different views, a probabilistic dependency relationship between views is established, and the performance of the learning model is finally enhanced.

[0082] The embodiment of the present application can model the multi-dimensional features of transaction points through the introduction of multi-view graphs, extract the correlation between different features, and improve the recognition accuracy of the model for short-term credit card behavior through graph representation learning.

[0083] See also Figure 4 As shown, in some implementations of the present application, based on the establishment of the node state sequence, two views are constructed according to the correlation between different aspects of the node states to achieve quantification of fine-grained credit card customer behavior information. At this time, the multi-view graph may include: a frequency graph and a time graph. The step of converting the node state sequence into a multi-view graph in S103 may specifically include:

[0084] S401: Based on the state transition frequencies between different discretized node states, convert the node state sequence into a frequency graph, wherein the nodes of the frequency graph represent the discretized node states, and the edges of the frequency graph represent the state transition frequencies between different discretized node states.

[0085] In an embodiment of the present application, a frequency graph f-Graph is used to measure the frequency of transitions between different node states of a specific customer during a short-term credit card usage period. In the frequency graph, nodes represent discretized node states, and edges represent the frequency of changes between different discretized node states. The more frequently a customer transitions between two node states, the more active the customer's specific consumption or repayment behavior is.

[0086] S402: Based on the average state transition time between different discretized node states, convert the node state sequence into a time graph, wherein the nodes of the time graph represent the discretized node states, and the edges of the time graph represent the average state transition time between different discretized node states.

[0087] In an embodiment of the present application, the conversion of the time graph is similar to the frequency graph, and the time graph is also related to the consumption or repayment behavior of a specific customer during a short-term credit card usage period. Specifically, within a given time period, the time required for each conversion from one discretized node state to another discretized node state and the number of times the node state conversion occurs can be counted, and then the average value is calculated as the average conversion time. Similarly, the nodes in the time graph represent the discretized node states, and the edges represent the average state conversion time between different discretized node states. Among them, the longer the average conversion time means that the customer's consumption or repayment behavior tends to be more stable; conversely, the shorter the average conversion time means that the customer's consumption or repayment behavior tends to be more sudden and hasty.

[0088] S104: Recognize credit card behavior features of unlabeled samples through the trained multi-view graph representation learning model.

[0089] In the embodiments of the present application, through feature recognition, accurate classification of customers' recent credit card usage behavior can be achieved, providing real-time support for the bank's risk management and personalized services. Compared with traditional feature engineering methods, graph representation learning can handle complex multidimensional features, improving recognition accuracy and applicability.

[0090] exist Figure 1 In the process shown, first, starting from the original transaction flow data, the transaction flow data is preprocessed. The preprocessed data samples include labeled samples with labels and unlabeled unlabeled samples. The credit card behavior feature type labels in the labeled samples can provide supervision information to build an effective model; then, the node state is calculated based on each transaction point of the labeled samples to obtain a node state sequence, so that the continuous transaction behavior is divided into a state transition process, thereby accurately reflecting the customer's credit card usage behavior characteristics from the perspective of short-term changes; then, the node state sequence is converted into a multi-view graph, and the multi-view graph representation learning model is trained through the multi-view graph, which can extract the characteristics of credit card usage behavior from different perspectives, so that the model can learn more comprehensive behavioral features instead of relying on a single dimension, so that it has a strong fine-grained analysis capability in time changes, which is suitable for short-term behavior recognition; finally, the unlabeled samples are feature recognized through the trained model, and the high expression ability of the multi-view graph representation learning model is used for inference to adapt to the rapidly changing behavioral characteristics in the short term. It can be seen that this application can accurately identify the characteristics of credit card transaction behavior in a short period of time through data preprocessing, node state calculation, node state sequence and multi-view graph construction, combined with a multi-view graph representation learning model.

[0091] In addition, based on credit card transaction flow data, this application first constructs graph data from multiple perspectives to quantify customers' complex and changeable credit card usage behavior information; then, based on graph representation learning that retains graph structure features, a multi-graph convolutional neural network is designed to extract implicit features of graph data from different perspectives; finally, using the trained network model, the customer's fine-grained credit card usage behavior characteristics can be predicted in a timely and relatively accurate manner. At the same time, this application also has good scalability, and can promote universal credit card transaction flow data to transaction flow data of other banking products such as funds and stocks, so as to realize the extraction and identification of other transaction behavior characteristics of customers, thereby helping to improve the overall competitiveness of banks. In addition, this application does not need to use the customer's sensitive information to analyze the behavioral characteristics of credit card use, but instead identifies based on desensitized credit card transaction flow data, which can better protect customer privacy and security.

[0092] In the above technical solution, for each data sample, two corresponding views can be constructed based on the multi-view graph data construction method. However, since the two views are heterogeneous, they can only characterize the characteristics of customer credit card transaction behavior from each isolated angle. Therefore, the recognition model based on a single view may produce incorrect or even contradictory recognition results. To this end, it is necessary to fuse multiple views into a complete and systematic whole.

[0093] To this end, the embodiment of the present application extracts and identifies fine-grained credit card customer behavior features based on multi-view graph representation learning. Compared with traditional convolutional neural networks, graph convolutional neural networks can effectively extract semantically rich features from complex and heterogeneous graph data. The multi-view graph representation learning model includes: a graph convolutional network and a classifier;

[0094] The graph convolutional network is used to extract and pool the node features of each node in the multi-view graph, and flatten the node features of each node into one dimension and map them to the sample label space to obtain the probability distribution of data samples on different credit card customer behavior feature categories; specifically, the graph convolutional network first extracts features from each node in the multi-view graph, and then after extracting the features of each node, these features are fused to integrate different features into a node representation with richer information, so that the information of each node is more comprehensive. After that, the node features are pooled to further integrate the features in the graph to reduce the complexity of the features while maintaining the key information in the features. After pooling, the graph convolutional network will flatten the pooled node features into a one-dimensional vector, that is, compress these features into a one-dimensional data sequence. The flattened feature form is conducive to the subsequent classification task processing. Finally, the fully connected layer is used to map the new feature representation of each node after flattening to the sample label space to obtain the probability distribution of data samples on different credit card customer behavior feature categories.

[0095] The classifier assigns a correct credit card customer behavior feature category label to a given data sample based on the feature representation extracted by the graph convolutional network. Specifically, for the input data sample, the softmax classifier is used to act on the feature vector extracted by the graph convolutional network to obtain the probability of it belonging to different credit card customer behavior feature categories. Furthermore, the final prediction result takes the credit card customer behavior feature category corresponding to the maximum probability.

[0096] It can be seen that the graph convolutional network completes the process from feature extraction, feature fusion to node classification in the entire process, so that a given credit card transaction flow segment can be accurately classified into a certain credit card customer behavior feature category, thereby helping to identify the customer's behavioral characteristics.

[0097] In an embodiment of the present application, the graph convolutional network includes four components, namely, a graph convolution layer, a graph embedding pooling layer, a flattening layer and a fully connected layer.

[0098] The graph convolution layer is used to extract the features of each node in the multi-view graph to obtain the initial feature representation of each node in the multi-view graph;

[0099] The graph embedding pooling layer pools the initial feature representation of each node through weighted convex combination to obtain a new feature representation of each node in the multi-view graph;

[0100] The flattening layer is used to flatten the new feature representation of each node after pooling into one dimension;

[0101] The fully connected layer is used to map the new feature representation of each flattened node to the sample label space to obtain the probability distribution of the data sample on different credit card customer behavior feature categories.

[0102] The design of each network layer and classifier will be introduced in detail below.

[0103] As for the specific implementation of the graph convolution layer, the graph convolution layer in the graph convolution network introduces filters from the perspective of graph signal processing to define the graph convolution operation, which can simultaneously perform feature learning from the attributes of the nodes and edges of the graph, and can realize learning from the adjacency matrix of multiple graphs with separate edge features. In the present invention, the input of the graph convolution layer is the multi-view graph data of the data sample, and the output is the feature representation of each node. . The embodiment of the present application uses a binary group G = (S, A) to represent the graph data corresponding to a single data sample, where S∈R N×2 Represents all possible node states, which means the position of each node in the original two-dimensional space; A∈R N×N×2 A tensor consisting of two matrices stacked together (A f ,A a ), where A f and A a They are the adjacency matrices corresponding to the frequency graph f-Graph and the time graph a-Graph, which directly reflect the relationship between each node and its adjacent nodes.

[0104] Based on the tensor A, the present embodiment introduces a convolution filter H∈R N×N×2 , which is the matrix H (1) , H (2) A stacked tensor. For each slice H of the tensor H (c) (c=1,2,), which is defined as follows:

[0105]

[0106] in and They correspond to the weights of the unit matrix and the two constructed adjacency matrices respectively; I is the unit matrix, which is used to extract and fuse the features of the node itself. According to formula (1), each element in the slice is a linear combination of the node itself and its two different types of neighboring nodes. In graph representation learning, these weights are randomly initialized and iteratively updated as the network is trained until the model converges. Each slice of the convolution filter H involves an attribute of the credit card transaction status, namely the credit limit and the remaining balance. On this basis, combined with the two attributes of the credit card transaction status and the relationship between the transaction status, the fusion feature of the node can be calculated according to the following formula:

[0107]

[0108] in is the learnable bias vector. Similarly, b1 is also iteratively updated during the learning process.

[0109] It is worth noting that in a graph convolution layer, a single filter often cannot fully capture the complex credit card customer behavior information contained in the multi-view graph data. Therefore, the embodiment of the present application constructs j-1 new convolution filters according to formula (1). Then, j-1 new features can be obtained according to formula (2). Finally, all j different V out ,Right now to represent the data sample. conv In the matrix, each row vector is a new rough representation of the node, and in the first graph convolutional layer it is the new representation of the corresponding credit card transaction state.

[0110] As for the specific implementation of the graph embedding pooling layer, this layer is used to extract V in a simplified space by reducing the number of nodes. conv A more compact feature representation. Different from the traditional pooling method that simply extracts compact representations from the neighborhood by averaging or maximum pooling, the embodiment of the present application uses a fully connected method to perform weighted convex combination of all node representations in the previous layer to obtain a new node representation. The weights between nodes are determined by a new convolution filter, which is constructed in the same way as the filter in the graph convolution layer.

[0111] For the output V of the previous network layer conv , an embedding column vector can be expressed as:

[0112]

[0113] in is the i-th slice of the filter, calculated by formula (1); Yes V conv The i-th column vector of ; b2 is the learnable bias column vector. Then, Vemb The softmax function is used to normalize so that the sum of the weights of all new nodes after the convolution filter operation is 1. The normalization process is as follows:

[0114] V norm =softmax(V emb ) (4)

[0115] Similarly, construct N′ different filters H emb To obtain a set of normalized embedding column vectors, and aggregate all column vectors to obtain a normalized embedding matrix, that is,

[0116] Furthermore, with the help of the normalized weights learned by each filter, that is, each column of V′, the new reduced-dimensional node representation It can be obtained by weighted combination of all node representations in the previous network layer, which can be calculated according to the following formula:

[0117] V pool =V ′T V conv (5)

[0118] Correspondingly, tensor A, that is, the two adjacency matrices corresponding to multiple views, should also be transformed and updated. Specifically, it can be calculated by the following formula:

[0119]

[0120] Among them A (l) Represents the lth slice of tensor A; is the new reduced dimension tensor It is worth noting that if another set of graph convolutional layers and graph pooling layers are required in the learning network, then the pooled graph data, i.e. A pool and V pool The sum will be used as input for further iterative processing.

[0121] As for the specific implementation of the flattening layer, in the flattening layer, the pooled nodes represent V pool , is flattened into a long column vector

[0122] As for the specific implementation of the fully connected layer, in order to map the feature vector to the sample label space, four neurons are constructed in the fully connected layer. The four neurons correspond to four different credit card customer behavior characteristic categories, including high-end customers, potential customers, mid-end customers and low-end customers, and are defined as follows:

[0123]

[0124] in represents the learned weight matrix; Represents the bias column vector of learning. In m', each neuron is used to calculate the probability that the data sample belongs to the corresponding credit card customer behavior feature category.

[0125] As for the specific implementation of the classifier, the embodiment of the present application uses a softmax classifier to assign a correct label to a given sample. Based on the graph convolutional neural network, for the input credit card transaction flow data sample T S , the feature representation m' extracted by the softmax classifier on the graph convolutional neural network can be obtained, and its true label belongs to the credit card customer behavior feature category d k The probability of

[0126] p, which is defined as follows:

[0127]

[0128] Where m ’(k) represents the output of the kth neuron in the fully connected layer of the credit card transaction flow data sample. Furthermore, the final prediction result (i.e. ) is the credit card customer behavior feature category corresponding to the maximum probability, which can be expressed as:

[0129]

[0130] The optimization goal of the recognition model based on graph convolutional networks is to calculate the classification loss of training samples through cross entropy, which can be formalized as:

[0131]

[0132] Where N l Indicates the number of training samples; The i-th training sample belongs to the d-th credit card customer behavior feature category k The predicted probability of The kth item of the one-hot encoding corresponding to the true label of the i-th training sample is obtained. Finally, in the iterative training process, the network parameters are continuously updated by using the error back propagation algorithm to obtain better network weights.

[0133] Next, through Figure 5 The process shown summarizes the solution provided in the embodiments of the present application.

[0134] Figure 5The figure shows a framework diagram of a fine-grained credit card customer behavior feature recognition system based on multi-view graph representation learning proposed in an embodiment of the present application, which is composed of three components: credit card transaction flow data preprocessing, multi-view graph data construction, and a credit card customer behavior feature recognition model based on multi-view graph representation learning. Among them, the credit card transaction flow data preprocessing component is mainly used to generate desensitized data fragments, i.e., data samples, from credit card transaction flow data. Multi-view graph data construction is used to quantify fine-grained credit card customer behavior information. The credit card customer behavior feature recognition model based on multi-view graph representation learning is used to effectively extract semantically rich features from complex and heterogeneous graph data, and classify and recognize the output high-level feature representation using a classifier.

[0135] In addition, in order to effectively identify the fine-grained credit card usage behavior characteristics of customers, the model needs to perform the following two processes: First, the graph convolutional network training process indicated by ① to ④ in the figure, the input is the labeled sample, and the output is the trained graph convolutional network. During the iterative training of the network, the labeled training samples are used to continuously adjust the network weights until the network converges. Second, the process of identifying fine-grained credit card usage behavior characteristics indicated by ① to ③ in the figure, the input is the credit card transaction flow data of a specific customer, and the output is the fine-grained credit card customer usage behavior characteristics of the customer. In other words, for a given credit card transaction flow segment, that is, an unlabeled sample, it can be assigned a correct label.

[0136] See also Figure 6 As shown, the embodiment of the present application provides a credit card behavior feature recognition device, the device comprising:

[0137] The data processing module 601 is used to obtain transaction flow data and pre-process the transaction flow data to obtain multiple data samples, wherein the data samples include marked samples and unmarked samples, and the marked samples are marked with credit card behavior feature type;

[0138] A sequence generation module 602 is used to generate a node state sequence corresponding to each marked sample by performing node state calculation on each marked sample;

[0139] A model training module 603 is used to convert the node state sequence into a multi-view graph, and train a multi-view graph representation learning model based on the multi-view graph;

[0140] The feature recognition module 604 is used to recognize the credit card behavior features of unlabeled samples through the trained multi-view graph representation learning model.

[0141] In some implementations of the embodiments of the present application, the transaction flow data is preprocessed to obtain multiple data samples, including:

[0142] Divide the transaction flow data into multiple continuous and equal-length data segments to obtain multiple data samples;

[0143] A plurality of data samples for model training are randomly selected from a plurality of data samples, and credit card behavior feature type tags are added to the plurality of data samples for model training to obtain a plurality of labeled samples.

[0144] In some implementations of the embodiments of the present application, the transaction flow data is composed of a series of transaction points sorted by time; the transaction flow data is divided into a plurality of continuous and equal-length data segments, including:

[0145] Starting from the first transaction point of the transaction flow data, through a time window of preset length, continuously slide to the last transaction point of the transaction flow data with a step of preset length to obtain multiple continuous and equal-length data fragments, and the data fragments are the same length as the time window.

[0146] In some implementations of the embodiments of the present application, by performing node state calculation on each marked sample, a node state sequence corresponding to each marked sample is generated, including:

[0147] For each marked sample, the node status of each transaction point in the marked sample is described by the credit limit and the remaining limit;

[0148] The credit limit range is divided into m credit limit discretization intervals, and the remaining limit range is divided into n remaining limit discretization intervals, and the m credit limit discretization intervals and the n remaining limit discretization intervals constitute m*n discretization node states, where m and n are both positive integers;

[0149] Based on the node state of each transaction point, each transaction point is mapped to the corresponding discretized node state to obtain the node state sequence corresponding to the labeled sample.

[0150] In some implementations of the embodiments of the present application, the multi-view graph includes: a frequency graph and a time graph, and converting the node state sequence into the multi-view graph includes:

[0151] Based on the state transition frequency between different discretized node states, the node state sequence is converted into a frequency graph, wherein the nodes of the frequency graph represent the discretized node states, and the edges of the frequency graph represent the state transition frequency between different discretized node states;

[0152] Based on the average state transition time between different discretized node states, the node state sequence is converted into a time graph, where the nodes of the time graph represent the discretized node states, and the edges of the time graph represent the average state transition time between different discretized node states.

[0153] In some implementations of the embodiments of the present application, the multi-view graph representation learning model includes: a graph convolutional network and a classifier;

[0154] The graph convolutional network is used to extract and pool the node features of each node in the multi-view graph, flatten the node features of each node into one dimension and map them to the sample label space to obtain the probability distribution of data samples on different credit card customer behavior feature categories;

[0155] The classifier assigns credit card customer behavior feature category labels to data samples based on the feature representation extracted by the graph convolutional network.

[0156] In some implementations of the embodiments of the present application, the graph convolution network includes: a graph convolution layer, a graph embedding pooling layer, a flattening layer, and a fully connected layer;

[0157] The graph convolution layer is used to extract the features of each node in the multi-view graph to obtain the initial feature representation of each node in the multi-view graph;

[0158] The graph embedding pooling layer pools the initial feature representation of each node through weighted convex combination to obtain a new feature representation of each node in the multi-view graph;

[0159] The flattening layer is used to flatten the new feature representation of each node after pooling into one dimension;

[0160] The fully connected layer is used to map the new feature representation of each flattened node to the sample label space to obtain the probability distribution of the data sample on different credit card customer behavior feature categories.

[0161] like Figure 7 As shown, an embodiment of the present application also provides a credit card behavior feature recognition device, including: a memory 6701 and a processor 702; wherein the memory 701 is used to store programs; the processor 702 is used to execute the programs in the memory to implement the various steps of the credit card behavior feature recognition device method provided in the embodiment of the present application.

[0162] The embodiment of the present application also provides a readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the various steps of the credit card behavior feature recognition device method provided in the embodiment of the present application are implemented.

[0163] Finally, it should be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements limited by the statement "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment including the elements.

[0164] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A credit card behavior feature recognition method, characterized in that: The method comprises: Acquire transaction flow data, and pre-process the transaction flow data to obtain multiple data samples, wherein the data samples include marked samples and unmarked samples, and the marked samples are marked with credit card behavior feature type; By performing node state calculation on each of the marked samples, a node state sequence corresponding to each of the marked samples is generated; Converting the node state sequence into a multi-view graph, and training a multi-view graph representation learning model based on the multi-view graph; The credit card behavior features of the unlabeled samples are identified through the trained multi-view graph representation learning model.

2. The method according to claim 1, characterized in that: The preprocessing of the transaction flow data to obtain a plurality of data samples includes: Dividing the transaction flow data into a plurality of continuous data segments of equal length to obtain the plurality of data samples; A plurality of data samples for model training are randomly selected from the plurality of data samples, and the credit card behavior feature type labels are added to the plurality of data samples for model training to obtain a plurality of labeled samples.

3. The method according to claim 2, characterized in that The transaction flow data is composed of a series of transaction points sorted by time; the transaction flow data is divided into a plurality of continuous and equal-length data segments, including: Starting from the first transaction point of the transaction flow data, through a time window of a preset length, continuously slide to the last transaction point of the transaction flow data with a step of the preset length to obtain the multiple continuous and equal-length data segments, and the data segments are the same as the length of the time window.

4. The method according to claim 3, characterized in that The step of performing node state calculation on each of the marked samples to generate a node state sequence corresponding to each of the marked samples includes: For each of the marked samples, describing the node status of each transaction point in the marked sample by credit limit and remaining limit; Divide the credit limit range into m credit limit discretization intervals, divide the remaining limit range into n remaining limit discretization intervals, and form m*n discretization node states by the m credit limit discretization intervals and the n remaining limit discretization intervals, wherein m and n are both positive integers; Based on the node state of each transaction point, each transaction point is mapped to a corresponding discretized node state to obtain the node state sequence corresponding to the marked sample.

5. The method according to claim 4, characterized in that The multi-view graph includes: a frequency graph and a time graph, and the step of converting the node state sequence into a multi-view graph includes: Based on the state transition frequency between different discretized node states, converting the node state sequence into a frequency graph, wherein the nodes of the frequency graph represent the discretized node states, and the edges of the frequency graph represent the state transition frequency between the different discretized node states; Based on the average state transition time between different discretized node states, the node state sequence is converted into a time graph, wherein the nodes of the time graph represent the discretized node states, and the edges of the time graph represent the average state transition time between the different discretized node states.

6. The method according to claim 1, characterized in that The multi-view graph representation learning model includes: a graph convolutional network and a classifier; The graph convolutional network is used to extract and pool node features of each node in the multi-view graph, and flatten the node features of each node into one dimension and then map them to the sample label space to obtain the probability distribution of data samples on different credit card customer behavior feature categories; The classifier assigns credit card customer behavior feature category labels to the data samples based on the feature representation extracted by the graph convolutional network.

7. The method according to claim 6, characterized in that The graph convolution network includes: a graph convolution layer, a graph embedding pooling layer, a flattening layer and a fully connected layer; The graph convolution layer is used to extract features of each node in the multi-view graph to obtain an initial feature representation of each node in the multi-view graph; The graph embedding pooling layer pools the initial feature representation of each node through weighted convex combination to obtain a new feature representation of each node in the multi-view graph; The flattening layer is used to flatten the new feature representation of each node after the pooling into one dimension; The fully connected layer is used to map the new feature representation of each flattened node to the sample label space to obtain the probability distribution of the data sample on different credit card customer behavior feature categories.

8. A credit card behavior feature recognition device, characterized in that: The device comprises: A data processing module, used to obtain transaction flow data, and pre-process the transaction flow data to obtain multiple data samples, wherein the data samples include marked samples and unmarked samples, and the marked samples are marked with credit card behavior feature type; A sequence generation module, used to generate a node state sequence corresponding to each of the marked samples by performing node state calculation on each of the marked samples; A model training module, used for converting the node state sequence into a multi-view graph, and training a multi-view graph representation learning model based on the multi-view graph; The feature recognition module is used to recognize the credit card behavior features of the unlabeled samples through the trained multi-view graph representation learning model.

9. A credit card behavior feature recognition device, characterized in that: The device includes a memory and a processor, and the processor is used to execute the program stored in the memory to run the credit card behavior feature recognition method as described in any one of claims 1 to 7.

10. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the credit card behavior feature recognition method as described in any one of claims 1 to 7 is implemented.

Images