Intelligent auxiliary credit granting method for consumer finance

Abstract

The invention discloses an intelligent auxiliary credit granting method for consumer finance, and relates to the technical field of monitoring analysis, and the method comprises the steps: constructing a customer initial feature set, carrying out the null value feature vector mapping analysis and intelligent completion of the customer initial feature set, generating a customer complete feature matrix, and carrying out the cross-scene model parameter migration of a risk prediction initial model. A risk prediction optimization model is obtained; performing real-time risk scanning on the customer complete feature matrix based on a risk prediction optimization model, generating a risk feature fluctuation map, extracting a key risk influence factor, performing risk conduction path simulation according to the key risk influence factor, and generating a dynamic risk early warning signal; performing credit extension decision tree reconstruction according to the dynamic risk early warning signal, generating an adaptive credit extension decision model, generating a gradient credit extension limit interval, performing decision weight fusion on the gradient credit extension limit interval and the dynamic risk early warning signal, and outputting an intelligent credit extension decision report. The method has the effect of improving the auxiliary credit extension efficiency.

Description

Technical Field

[0001] The present application relates to the field of monitoring and analysis technology, and in particular to a method for intelligent assisted credit granting in consumer finance. Background Art

[0002] Consumers of financial products are highly concerned with the sense of gain from using these products, such as credit limits. They expect consumer finance products to contribute to a better life, including enabling flexible and even pre-consumption through higher credit limits, and obtaining greater benefits through promotional offers. Furthermore, consumers are concerned about the product experience, including product approval speed, ease of use, and the quality of the institution's service. Reasonable and reliable credit limits can enhance users' sense of gain. This is directly related to consumer stickiness and loyalty. Currently, credit approval in the consumer finance sector generally faces the dual challenges of data integrity and timeliness of risk assessment.

[0003] In related technologies, traditional credit granting methods mainly rely on manual collection of structured data for credit assessment. For the common problem of missing features in multi-source heterogeneous data, fixed rules are usually used to fill in the gaps or return them for manual processing, resulting in inefficient data preprocessing and significant subjective bias. It is also impossible to identify potential risks early in the process, resulting in a waste of resources. Auditors need to infer missing data or risk points on their own, which reduces the efficiency of auxiliary credit granting and leaves room for improvement. Summary of the Invention

[0004] In response to the deficiencies of the existing technology, this application provides a consumer finance intelligent assisted credit granting method.

[0005] In a first aspect, the present application provides a method for intelligently assisting credit granting for consumer finance, comprising the following steps:

[0006] Step D1: Obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, construct the customer's initial feature set, perform null value feature vector mapping analysis on the customer's initial feature set, generate a null value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null value feature distribution heat map to generate a complete customer feature matrix;

[0007] Step D2: Reorganize the time-varying feature sequence of the customer's complete feature matrix to construct a dynamic risk feature timeline. Based on the dynamic risk feature timeline, construct a pre-risk dynamic analysis model to generate an initial risk prediction model. Migrate the model parameters across scenarios to obtain a risk prediction optimization model.

[0008] Step D3: Based on the risk prediction optimization model, perform real-time risk scanning on the customer's complete feature matrix, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate risk transmission paths based on the key risk influencing factors, and generate dynamic risk warning signals;

[0009] Step D4: Reconstruct the credit decision tree based on the dynamic risk warning signal to generate an adaptive credit decision model. Use the adaptive credit decision model to calculate the credit limit elasticity of the customer's complete feature matrix, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal, and output an intelligent credit decision report.

[0010] Preferably, step D1 comprises the following steps:

[0011] Step D11: Connect to the API interface of relevant data sources and build a multi-source heterogeneous data collection interface;

[0012] Step D12: Collect and obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, and perform vectorization processing on the unstructured data to generate the customer's initial feature set;

[0013] Step D13: Use a feature null value detection algorithm to perform null value feature vector mapping analysis on the customer's initial feature set to generate a null value feature distribution heat map;

[0014] Step D14: performing null value distribution pattern recognition on the null value feature distribution heat map to construct a null value feature association map;

[0015] Step D15: Use the spatiotemporal attention mechanism to complete the contextual features of the null-value feature association map, and use the LSTM neural network to perform time series feature prediction and filling to generate a complete customer feature matrix.

[0016] Preferably, step D2 comprises the following steps:

[0017] Step D21: Segment the customer's complete feature matrix into sliding time windows to construct a time-varying feature sequence set;

[0018] Step D22: Using a dynamic time warping algorithm to perform time dimension alignment processing on the time-varying feature sequence set to generate a dynamic risk feature time axis;

[0019] Step D23: constructing a pre-risk dynamic analysis model based on the dynamic risk feature timeline to generate an initial risk prediction model;

[0020] Step D24: Based on the transfer learning framework, the credit assessment model parameters in the source domain are transferred to the target domain, and the domain adaptation network is used to align the feature distribution;

[0021] Step D25: Optimize the risk prediction model through adversarial training strategy and generate a deep neural network model, and then use the Bayesian optimization algorithm to automatically tune the hyperparameters of the deep neural network model to obtain the risk prediction optimization model.

[0022] Preferably, step D3 comprises the following steps:

[0023] Step D31: Input the customer's complete feature matrix into the risk prediction optimization model to obtain the risk probability distribution at each time point;

[0024] Step D32: Decompose the time-frequency characteristics of the risk probability distribution using a wavelet transform algorithm to extract the main frequency components of the risk fluctuations, and then construct a risk characteristic fluctuation map based on the main frequency components of the risk fluctuations;

[0025] Step D33: Using an independent component analysis algorithm to perform multi-dimensional risk indicator decoupling analysis on the risk characteristic fluctuation map, thereby isolating key risk influencing factors, which include income fluctuation factors, debt expansion factors, and behavioral abnormality factors;

[0026] Step D34: Construct a risk transmission network based on the system dynamics model to simulate the risk transmission paths of key risk factors in the capital chain, credit chain, and supply chain;

[0027] Step D35: Perform sensitivity analysis through the risk transmission path to determine the warning threshold, and then generate a dynamic risk warning signal based on the warning threshold.

[0028] Preferably, step D4 comprises the following steps:

[0029] Step D41: Adjust the branch weights of the credit decision tree based on the dynamic risk warning signal to build an adaptive decision model;

[0030] Step D42: Stress-test the customer's repayment ability through the adaptive decision-making model to generate credit limit ranges under different economic scenarios;

[0031] Step D43: Using fuzzy mathematics methods to calculate the membership of credit limit intervals under different economic scenarios, and determine the gradient credit limit interval;

[0032] Step D44: Construct a dynamic adjustment matrix for decision weights, perform nonlinear mapping between dynamic risk warning signals and gradient credit limit intervals, and balance the risk-return ratio through a multi-objective optimization algorithm to generate an intelligent credit decision report.

[0033] Preferably, the multi-objective optimization algorithm includes:

[0034] Define the optimization objective function set, optimize the objective function set, generate a set of candidate solutions, sort the candidate solutions, identify the optimal compromise solution, apply robust optimization theory to perform sensitivity analysis on the optimal compromise solution, and generate an intelligent credit decision report based on the results of the sensitivity analysis.

[0035] In a second aspect, the present application provides a consumer finance intelligent assisted credit granting system, comprising:

[0036] The data acquisition and analysis module is used to obtain basic customer data, financial transaction flow data, and third-party credit data through multi-source heterogeneous data acquisition interfaces, construct an initial customer feature set, perform null-value feature vector mapping analysis on the initial customer feature set, generate a null-value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null-value feature distribution heat map to generate a complete customer feature matrix;

[0037] The risk prediction optimization module is used to reorganize the time-varying feature sequence of the customer's complete feature matrix, construct a dynamic risk feature timeline, build a pre-risk dynamic analysis model based on the dynamic risk feature timeline, generate an initial risk prediction model, and migrate the model parameters of the initial risk prediction model across scenarios to obtain a risk prediction optimization model;

[0038] The early warning module is used to perform real-time risk scanning of the customer's complete feature matrix based on the risk prediction optimization model, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate the risk transmission path based on the key risk influencing factors, and generate dynamic risk early warning signals;

[0039] The report generation module is used to reconstruct the credit decision tree based on dynamic risk warning signals, generate an adaptive credit decision model, calculate the credit limit elasticity of the customer's complete feature matrix through the adaptive credit decision model, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal for decision weight, and output an intelligent credit decision report.

[0040] In a third aspect, the present application provides a computer-readable storage medium storing instructions. When the instructions are executed on a computer, the computer executes any one of the above-mentioned intelligent assisted credit granting methods for consumer finance.

[0041] In summary, this application has the following beneficial technical effects:

[0042] The present application provides a method for intelligent assisted credit granting in consumer finance, which collects and analyzes basic data, financial transaction flow data, and third-party credit data corresponding to a customer to generate a complete customer feature matrix, reorganizes the time-varying feature sequence of the customer complete feature matrix, constructs a dynamic risk feature timeline, constructs a front-end risk dynamic analysis model based on the dynamic risk feature timeline, obtains a risk prediction optimization model, performs real-time risk scanning on the customer complete feature matrix based on the risk prediction optimization model, generates a risk feature fluctuation map, performs multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extracts key risk influencing factors, simulates risk transmission paths based on the key risk influencing factors, and generates dynamic risk warning signals, thereby reducing the occurrence of low data preprocessing efficiency and significant subjective bias, and reconstructs a credit decision tree based on the dynamic risk warning signals to generate an adaptive credit decision model, performs credit limit elasticity calculation on the customer complete feature matrix, generates a gradient credit limit range, fuses the gradient credit limit range with the dynamic risk warning signal for decision weight, and outputs an intelligent credit decision report, thereby effectively identifying potential risks, reducing resource waste, and effectively improving assisted credit efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 This is a flow chart of the method for intelligent assisted credit granting of consumer finance in an embodiment of the present application.

[0045] Figure 2 This is a system diagram of the intelligent assisted credit granting of consumer finance in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following is combined with Figure 1-2 This application is described in further detail.

[0047] Example 1

[0048] The embodiments of the present application disclose a method for intelligent assisted credit granting in consumer finance.

[0049] Reference Figure 1 , a consumer finance intelligent assisted credit granting method, comprising the following steps:

[0050] Step D1: Obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, construct the customer's initial feature set, perform null value feature vector mapping analysis on the customer's initial feature set, generate a null value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null value feature distribution heat map to generate a complete customer feature matrix;

[0051] Step D2: Reorganize the time-varying feature sequence of the customer's complete feature matrix to construct a dynamic risk feature timeline. Based on the dynamic risk feature timeline, construct a pre-risk dynamic analysis model to generate an initial risk prediction model. Migrate the model parameters across scenarios to obtain a risk prediction optimization model.

[0052] Step D3: Based on the risk prediction optimization model, perform real-time risk scanning on the customer's complete feature matrix, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate risk transmission paths based on the key risk influencing factors, and generate dynamic risk warning signals;

[0053] Step D4: Reconstruct the credit decision tree based on the dynamic risk warning signal to generate an adaptive credit decision model. Use the adaptive credit decision model to calculate the credit limit elasticity of the customer's complete feature matrix, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal, and output an intelligent credit decision report.

[0054] It should be noted that step D1 includes the following steps:

[0055] Step D11: Connect to the API interface of relevant data sources and build a multi-source heterogeneous data collection interface;

[0056] Step D12: Collect and obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, and perform vectorization processing on the unstructured data to generate the customer's initial feature set;

[0057] Step D13: Use a feature null value detection algorithm to perform null value feature vector mapping analysis on the customer's initial feature set to generate a null value feature distribution heat map;

[0058] Step D14: performing null value distribution pattern recognition on the null value feature distribution heat map to construct a null value feature association map;

[0059] Step D15: Use the spatiotemporal attention mechanism to complete the contextual features of the null-value feature association map, and use the LSTM neural network to perform time series feature prediction and filling to generate a complete customer feature matrix.

[0060] Specifically, step D11: connect to the API interface of relevant data sources and build a multi-source heterogeneous data collection interface. In actual business, customer data comes from a wide range of sources and in various forms, which may include the company's internal customer management system, financial transaction system, and external third-party credit reporting agencies. In order to obtain data from different sources, it is necessary to connect to the API interface of each data source. For example, connect to the API interface of the customer management system to obtain basic customer information such as name, age, contact information, etc.; connect to the API interface of the financial transaction system to collect customer financial transaction flow data, including transaction time, transaction amount, transaction type, etc.; connect to the API interface of the third-party credit reporting agency to obtain customer third-party credit data such as credit score, credit record, etc. By integrating and encapsulating the above API interfaces, a unified multi-source heterogeneous data collection interface is constructed, which makes it possible to conveniently and efficiently collect data from different data sources;

[0061] Step D12: Collect and obtain the basic data, financial transaction flow data, and third-party credit data corresponding to the customer through a multi-source heterogeneous data acquisition interface, and perform vectorization processing on the unstructured data to generate an initial feature set for the customer. Utilize the multi-source heterogeneous data acquisition interface constructed in step D11 to collect customer data from various data sources according to predetermined rules and frequencies. The collected data may contain structured data (such as transaction amounts in tabular form) and unstructured data (such as the customer's credit description text). For unstructured data, vectorization processing is required to convert it into a numerical form that can be processed by a computer. For example, for the customer's credit description text, a word vector model in natural language processing (such as Word2Vec) can be used to map each word in the text into a vector. Then, the entire text is converted into a vector representation through a certain method (such as average vectorization). The processed structured data and the vector representation of the unstructured data are integrated to generate an initial feature set for the customer.

[0062] Step D13: Use the feature null value detection algorithm to perform null value feature vector mapping analysis on the customer's initial feature set to generate a null value feature distribution heat map. There may be some null values (missing values) in the customer's initial feature set. These null values will affect subsequent data analysis and model training. Use the feature null value detection algorithm to detect each feature in the feature set to determine whether it has null values. For features with null values, perform mapping analysis on their corresponding feature vectors to determine the position and distribution of the null values in the feature vector. For example, there are 3 null values in a feature vector, located at the 5th, 10th and 15th positions respectively. Based on the distribution of the above null values, generate a null value feature distribution heat map. In the heat map, different colors are used to represent the density of null values. The darker the color, the more null values there are, thereby intuitively showing the distribution of null values in the feature set.

[0063] Step D14: Perform null value distribution pattern recognition on the null value feature distribution heat map, construct a null value feature association map, and identify the distribution pattern of null values by observing the null value feature distribution heat map. For example, it is found that the null values of certain features always appear at the same time, or the null values of certain features are correlated with the values of other features. According to the above pattern, a null value feature association map is constructed. In the map, the features are used as nodes, and the correlation between null values is used as edges to show the null value correlation between features. For example, the null values of feature A and feature B often appear at the same time, so an edge connecting feature A and feature B is established in the map, and the null value correlation between them is marked;

[0064] Step D15: Use the spatiotemporal attention mechanism to complete the contextual features of the null-value feature association graph, and use the LSTM neural network to predict and fill in the time series features to generate a complete customer feature matrix. The spatiotemporal attention mechanism can focus on the contextual features related to the null-value features based on the correlation between the features in the null-value feature association graph, and extract useful information from the context to complete the null values. For example, if the null value of feature A is related to the values of feature B and feature C, the spatiotemporal attention mechanism will pay attention to the values of feature B and feature C, and use their information to infer the null value of feature A. At the same time, for features involving time series (such as transaction time in financial transaction flow data), the LSTM neural network is used to predict and fill in the time series features. The LSTM neural network can learn the long-term dependencies in time series data and predict the value of null values based on the existing time series data. The completed and filled features are integrated to generate a complete customer feature matrix. The complete customer feature matrix contains all the features of the customer and does not have null values, providing a high-quality data foundation for subsequent customer analysis and modeling;

[0065] It should be noted that step D2 includes the following steps:

[0066] Step D21: Segment the customer's complete feature matrix into sliding time windows to construct a time-varying feature sequence set;

[0067] Step D22: Using a dynamic time warping algorithm to perform time dimension alignment processing on the time-varying feature sequence set to generate a dynamic risk feature time axis;

[0068] Step D23: constructing a pre-risk dynamic analysis model based on the dynamic risk feature timeline to generate an initial risk prediction model;

[0069] Step D24: Based on the transfer learning framework, the credit assessment model parameters in the source domain are transferred to the target domain, and the domain adaptation network is used to align the feature distribution;

[0070] Step D25: Optimize the risk prediction model through adversarial training strategy and generate a deep neural network model, and then use the Bayesian optimization algorithm to automatically tune the hyperparameters of the deep neural network model to obtain the risk prediction optimization model.

[0071] Specifically, step D21: perform a sliding time window segmentation on the customer's complete feature matrix to construct a time-varying feature sequence set. The customer's complete feature matrix contains various attributes and behavioral feature data of the customer. In order to capture the changes in customer characteristics over time, the sliding time window method is used to segment the matrix. Set a time window of fixed size, for example, one month as a time window, and let the time window slide sequentially on the time axis of the customer data. In each time window, extract the customer's feature data to form a feature sequence. For example, for the customer's financial transaction flow data, in the first time window, extract the transaction amount, transaction frequency and other features within the month; in the second time window, extract the corresponding features in the same way. Through the above method, the customer's complete feature matrix is divided into multiple time-varying feature sequences. Multiple time-varying feature sequences reflect the changes in customer characteristics in different time periods and construct a time-varying feature sequence set.

[0072] Step D22: Use the dynamic time warping algorithm to align the time dimension of the time-varying feature sequence set to generate a dynamic risk feature timeline. Since the feature sequences of different customers may differ in time length and time rhythm, in order to accurately compare and analyze the time-varying feature sequences, the dynamic time warping algorithm (DTW) is used. The core idea of the DTW algorithm is to find the best matching path between two time series by performing nonlinear alignment on the time axis. For example, the time length of customer A's transaction feature sequence is 10 time units, and the time length of customer B's transaction feature sequence is 12 time units, and their transaction rhythms are also different. The DTW algorithm will adjust the two sequences so that they can be aligned in the time dimension and find the similarity between them. This process is performed on all sequences in the time-varying feature sequence set. After they are aligned in the time dimension, a dynamic risk feature timeline is generated. On this timeline, the feature sequences of different customers are comparable in time, which facilitates subsequent analysis.

[0073] Step D23: Construct a pre-risk dynamic analysis model based on the dynamic risk feature timeline to generate an initial risk prediction model. After obtaining the dynamic risk feature timeline, construct a pre-risk dynamic analysis model based on the feature sequence on the timeline and the customer's risk history data (such as whether there has been overdue payments, defaults, etc.). Various algorithms in machine learning can be used, such as decision trees, support vector machines, or neural networks. For example, a neural network algorithm is used to train the model with the feature sequence on the dynamic risk feature timeline as input and the customer's risk label (such as risk level) as output. Through training, the pre-risk dynamic analysis model learns the relationship between the feature sequence and the risk, and generates an initial risk prediction model. The initial model can make a preliminary prediction of the customer's risk based on the customer's feature sequence.

[0074] Step D24: Based on the transfer learning framework, the credit assessment model parameters in the source domain are transferred to the target domain, and the domain adaptation network is used to align the feature distribution. Transfer learning is a method of transferring knowledge learned in one domain (source domain) to another domain (target domain). In customer risk prediction, there may be a source domain (such as a mature market with a large amount of credit assessment data) and a target domain (such as a newly developed market or a specific customer group). The parameters of the credit assessment model that has been trained in the source domain are transferred to the model in the target domain to speed up the training speed of the target domain model and improve its performance. At the same time, since the data feature distribution of the source domain and the target domain may be different, the domain adaptation network is used to align the feature distribution. The domain adaptation network adjusts the feature distribution to make the feature distribution of the target domain more similar to that of the source domain, thereby improving the generalization ability of the model in the target domain.

[0075] Step D25: Optimize the risk prediction model through adversarial training strategy and generate a deep neural network model, and then use the Bayesian optimization algorithm to automatically tune the hyperparameters of the deep neural network model to obtain a risk prediction optimization model. The adversarial training strategy is a method to enhance the robustness of the model by introducing adversarial samples (i.e., samples deliberately constructed to mislead the model). In the risk prediction model, the adversarial training strategy is used to optimize the model so that the risk prediction model can better cope with various complex situations and noise. After adversarial training, a deep neural network model is generated, and the Bayesian optimization algorithm is used to automatically tune the hyperparameters (such as learning rate, number of hidden layer nodes, etc.) of the deep neural network model. The Bayesian optimization algorithm continuously tries different hyperparameter combinations and adjusts the search direction according to the performance feedback of the model to find the optimal hyperparameter combination. After tuning, the risk prediction optimization model is obtained. The risk prediction optimization model has higher accuracy and reliability in customer risk prediction.

[0076] It should be noted that step D3 includes the following steps:

[0077] Step D31: Input the customer's complete feature matrix into the risk prediction optimization model to obtain the risk probability distribution at each time point;

[0078] Step D32: Decompose the time-frequency characteristics of the risk probability distribution using a wavelet transform algorithm to extract the main frequency components of the risk fluctuations, and then construct a risk characteristic fluctuation map based on the main frequency components of the risk fluctuations;

[0079] Step D33: Using an independent component analysis algorithm to perform multi-dimensional risk indicator decoupling analysis on the risk characteristic fluctuation map, thereby isolating key risk influencing factors, which include income fluctuation factors, debt expansion factors, and behavioral abnormality factors;

[0080] Step D34: Construct a risk transmission network based on the system dynamics model to simulate the risk transmission paths of key risk factors in the capital chain, credit chain, and supply chain;

[0081] Step D35: Perform sensitivity analysis through the risk transmission path to determine the warning threshold, and then generate a dynamic risk warning signal based on the warning threshold.

[0082] Specifically, step D31: input the customer's complete feature matrix into the risk prediction optimization model to obtain the risk probability distribution of each time node. The customer's complete feature matrix obtained through the above processing is input into the optimized risk prediction model. The risk prediction model is based on previous training and learning and can calculate the probability of the customer facing different risk levels at each time node according to the input feature data. For example, the risk prediction model can output that the probability of the customer being at a low risk level in the next month is 0.6, the probability of being at a medium risk level is 0.3, and the probability of being at a high risk level is 0.1. In this way, the risk probability distribution of each time node is obtained, providing basic data for subsequent risk analysis;

[0083] Step D32: Decompose the time-frequency characteristics of the risk probability distribution through the wavelet transform algorithm, extract the main frequency component of the risk fluctuation, and then construct a risk characteristic fluctuation spectrum based on the main frequency component of the risk fluctuation. The wavelet transform algorithm is a tool that can analyze time series data in two dimensions: time and frequency. The time series data of the risk probability distribution obtained in step D31 is processed using the wavelet transform algorithm. The algorithm can decompose the risk probability distribution into components of different frequencies, similar to decomposing a complex signal into a superposition of multiple simple sine waves. Among the above frequency components, the main frequency component with the largest energy share is extracted, which reflects the main trend and characteristics of the risk fluctuation. For example, after wavelet transform, it is found that the frequency corresponding to the main frequency component in the risk probability distribution is 0.1Hz, which means that the risk fluctuation has a main change trend with a period of 10 seconds. Based on this main frequency component and other relevant information, a risk characteristic fluctuation spectrum is constructed. The spectrum can use different colors, lines, etc. to intuitively display the characteristics and changes of risk fluctuations;

[0084] Step D33: Use the independent component analysis algorithm to perform multidimensional risk indicator decoupling analysis on the risk characteristic fluctuation map, and then separate the key risk influencing factors. The key risk influencing factors include income fluctuation factor, debt expansion factor and behavioral anomaly factor. The independent component analysis algorithm (ICA) can decompose multiple interrelated variables into independent components. The ICA algorithm is used to process the multidimensional risk indicator data contained in the risk characteristic fluctuation map. For example, risk indicators may include multiple aspects such as customer income changes, debt level, transaction frequency and amount. The ICA algorithm uses calculation and analysis to unravel the complex correlations between these indicators and separate independent key risk influencing factors. For example, analysis found that the income fluctuation factor reflects the instability of customer income. When income fluctuates greatly, the customer's repayment ability may be affected, thereby increasing risk; the debt expansion factor reflects the growth rate of customer debt. Excessive debt expansion will lead to increased debt repayment pressure; the behavioral anomaly factor covers abnormal performance of customers in financial transactions, such as sudden and frequent large-scale transactions. Such abnormal behavior may indicate potential risks.

[0085] Step D34: Construct a risk transmission network based on the system dynamics model to simulate the risk transmission paths of key risk factors in the capital chain, credit chain and supply chain. The system dynamics model is a model used to study the interactions and dynamic changes between various elements in a complex system. In customer risk assessment, a risk transmission network is constructed with key risk influencing factors as nodes and capital chain, credit chain and supply chain relationships as connecting edges. For example, in the capital chain, the customer's income volatility factor may affect their repayment ability, thereby affecting the financial status of the financial institution with which they have a lending relationship; in the credit chain, the customer's debt expansion factor may cause their credit rating to decline, thereby affecting the credit relationship with other partners; in the supply chain, the customer's abnormal behavior factor may affect the stability of the supply chain. Use the system dynamics model to simulate the risk transmission path and analyze the propagation and influence mechanism of key risk factors in different chains;

[0086] Step D35: Conduct sensitivity analysis through the risk transmission path to determine the warning threshold, and then generate a dynamic risk warning signal based on the warning threshold. Sensitivity analysis is to study the extent to which a certain factor affects the result when it changes. For the risk transmission path, analyze the extent to which changes in key risk factors affect the final risk result. For example, when the income volatility factor changes by a certain amplitude, how much will the customer's risk level change accordingly. Based on the results of the sensitivity analysis, determine a reasonable warning threshold. For example, when the income fluctuation exceeds a certain proportion, or the debt expansion reaches a certain level, it is considered that the risk has reached a level that requires a warning. Based on the warning threshold, when the monitored key risk factor data reaches or exceeds the threshold, a dynamic risk warning signal is generated to promptly notify relevant personnel to take measures to deal with potential risks;

[0087] It should be noted that step D4 includes the following steps:

[0088] Step D41: Adjust the branch weights of the credit decision tree based on the dynamic risk warning signal to build an adaptive decision model;

[0089] Step D42: Stress-test the customer's repayment ability through the adaptive decision-making model to generate credit limit ranges under different economic scenarios;

[0090] Step D43: Using fuzzy mathematics methods to calculate the membership of credit limit intervals under different economic scenarios, and determine the gradient credit limit interval;

[0091] Step D44: Construct a dynamic adjustment matrix for decision weights, perform nonlinear mapping between dynamic risk warning signals and gradient credit limit intervals, and balance the risk-return ratio through a multi-objective optimization algorithm to generate an intelligent credit decision report.

[0092] Specifically, step D41: Based on the dynamic risk warning signal, the branch weights of the credit decision tree are adjusted to construct an adaptive decision model. The credit decision tree is a model commonly used for credit decision-making. By analyzing and judging various characteristics of the customer, it is decided whether to grant credit and the amount of credit, etc. The dynamic risk warning signal reflects the current risk situation faced by the customer. After receiving the dynamic risk warning signal, the branch weights of the credit decision tree are adjusted according to the risk level indicated by the signal. For example, if the dynamic risk warning signal shows that the customer's risk level has increased, then in the decision tree, the branch weights related to high risk will increase accordingly, so that the adaptive decision model pays more attention to these high-risk factors in subsequent decisions. Assume that in the original decision tree, the branch weight for the feature of unstable income of the customer is 0.3. When a high-risk warning signal is received and the signal is related to the factor of unstable income, the branch weight is adjusted to 0.5. Through the above adjustment, an adaptive decision model that can automatically adapt to changes in risk is constructed, so that it can make more reasonable decisions under different risk situations;

[0093] Step D42: Use the adaptive decision model to stress test the customer's repayment ability and generate credit limit ranges under different economic scenarios. Use the adaptive decision model constructed in step D41 to conduct a comprehensive stress test on the customer's repayment ability. Set a variety of different economic scenarios, such as economic prosperity, economic recession, industry fluctuations, etc. In each economic scenario, consider the customer's various sources of income, debt situation, expenditure items and other factors to simulate the customer's repayment ability under different economic environments. For example, in an economic recession scenario, assuming that the customer's industry is hit hard, the income may drop by 30%, while the repayment pressure of its debt remains unchanged. Through the adaptive decision model, a comprehensive analysis of the above factors is conducted to calculate the maximum and minimum credit limits that the customer can bear in this economic scenario, thereby generating credit limit ranges under different economic scenarios. For example, in an economic prosperity scenario, the credit limit range may be 100,000 to 200,000 yuan; in an economic recession scenario, the credit limit range may become 50,000 to 100,000 yuan;

[0094] Step D43: Use fuzzy mathematics methods to calculate the membership of the credit limit intervals under different economic scenarios and determine the gradient credit limit intervals. Fuzzy mathematics methods can handle problems of uncertainty and ambiguity. For the credit limit intervals under different economic scenarios obtained in step D42, use fuzzy mathematics methods to calculate the membership of each credit limit value under different economic scenarios. For example, for a credit limit of 160,000 yuan, its membership in the economic prosperity scenario is calculated to be 0.6, and its membership in the economic recession scenario is 0.3. According to the above membership, the credit limit intervals under different economic scenarios are integrated to determine the gradient credit limit interval. The gradient credit limit interval takes into account the possibility and uncertainty of the credit limit under different economic scenarios, and divides the credit limit into different gradient ranges. For example, the credit limit is divided into a low-risk gradient (below 50,000 yuan), a medium-low risk gradient (50,000 to 100,000 yuan), a medium-risk gradient (100,000 to 200,000 yuan), etc. Each gradient has a corresponding membership and risk level;

[0095] Step D44: Construct a dynamic adjustment matrix of decision weights, perform nonlinear mapping between dynamic risk warning signals and gradient credit limit intervals, and balance the risk-return ratio through a multi-objective optimization algorithm to generate an intelligent credit decision report. First, construct a dynamic adjustment matrix of decision weights, which is used to describe the relationship between dynamic risk warning signals and gradient credit limit intervals. Perform nonlinear mapping between different levels of dynamic risk warning signals and different gradients of gradient credit limit intervals, that is, determine the corresponding credit limit interval based on the strength of the risk warning signal. For example, when the dynamic risk warning signal is high risk, the corresponding credit limit interval may be limited to a low-risk gradient range. Then, through a multi-objective optimization algorithm, the credit decision is optimized under the premise of considering risk and return. The multi-objective optimization algorithm balances the relationship between risk and return, so that credit decisions can both control risks and achieve certain returns. For example, when the risk is high, the credit limit can be appropriately reduced to reduce potential losses; when the risk is low, the credit limit can be increased to obtain more returns. An intelligent credit decision report is generated based on the optimization results. The report contains the customer's basic information, risk assessment results, gradient credit limit range, and decision-making basis, providing a comprehensive and scientific reference for credit decisions.

[0096] Furthermore, the multi-objective optimization algorithm includes:

[0097] Define the optimization objective function set, optimize the objective function set, generate a set of candidate solutions, sort the candidate solutions, identify the optimal compromise solution, apply robust optimization theory to perform sensitivity analysis on the optimal compromise solution, and generate an intelligent credit decision report based on the results of the sensitivity analysis.

[0098] Example 2

[0099] The embodiment of the present application also discloses a consumer finance intelligent assisted credit system.

[0100] Reference Figure 2 , a consumer finance intelligent auxiliary credit system, comprising:

[0101] The data acquisition and analysis module is used to obtain basic customer data, financial transaction flow data, and third-party credit data through multi-source heterogeneous data acquisition interfaces, construct an initial customer feature set, perform null-value feature vector mapping analysis on the initial customer feature set, generate a null-value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null-value feature distribution heat map to generate a complete customer feature matrix;

[0102] The risk prediction optimization module is used to reorganize the time-varying feature sequence of the customer's complete feature matrix, construct a dynamic risk feature timeline, build a pre-risk dynamic analysis model based on the dynamic risk feature timeline, generate an initial risk prediction model, and migrate the model parameters of the initial risk prediction model across scenarios to obtain a risk prediction optimization model;

[0103] The early warning module is used to perform real-time risk scanning of the customer's complete feature matrix based on the risk prediction optimization model, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate the risk transmission path based on the key risk influencing factors, and generate dynamic risk early warning signals;

[0104] The report generation module is used to reconstruct the credit decision tree based on dynamic risk warning signals, generate an adaptive credit decision model, calculate the credit limit elasticity of the customer's complete feature matrix through the adaptive credit decision model, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal for decision weight, and output an intelligent credit decision report.

[0105] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in similar ways. As long as they do not deviate from the concept of the invention, they should all fall within the scope of protection of the present invention.

[0106] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0107] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention.

Claims

1. A consumer finance intelligent assisted credit granting method, characterized in that: The following steps are involved: Step D1: Obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, construct the customer's initial feature set, perform null value feature vector mapping analysis on the customer's initial feature set, generate a null value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null value feature distribution heat map to generate a complete customer feature matrix; Step D2: Reorganize the time-varying feature sequence of the customer's complete feature matrix to construct a dynamic risk feature timeline. Based on the dynamic risk feature timeline, construct a pre-risk dynamic analysis model to generate an initial risk prediction model. Migrate the model parameters across scenarios to obtain a risk prediction optimization model. Step D3: Based on the risk prediction optimization model, perform real-time risk scanning on the customer's complete feature matrix, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate risk transmission paths based on the key risk influencing factors, and generate dynamic risk warning signals; Step D4: Reconstruct the credit decision tree based on the dynamic risk warning signal to generate an adaptive credit decision model. Use the adaptive credit decision model to calculate the credit limit elasticity of the customer's complete feature matrix, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal, and output an intelligent credit decision report.

2. A consumer finance intelligent assisted credit granting method according to claim 1, characterized in that: Step D1 includes the following steps: Step D11: Connect to the API interface of relevant data sources and build a multi-source heterogeneous data collection interface; Step D12: Collect and obtain the customer's corresponding basic data, financial transaction flow data, and third-party credit data through a multi-source heterogeneous data acquisition interface, and perform vectorization processing on the unstructured data to generate the customer's initial feature set; Step D13: Use a feature null value detection algorithm to perform null value feature vector mapping analysis on the customer's initial feature set to generate a null value feature distribution heat map; Step D14: performing null value distribution pattern recognition on the null value feature distribution heat map to construct a null value feature association map; Step D15: Use the spatiotemporal attention mechanism to complete the contextual features of the null-value feature association map, and use the LSTM neural network to perform time series feature prediction and filling to generate a complete customer feature matrix.

3. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that: Step D2 includes the following steps: Step D21: Segment the customer's complete feature matrix into sliding time windows to construct a time-varying feature sequence set; Step D22: Using a dynamic time warping algorithm to perform time dimension alignment processing on the time-varying feature sequence set to generate a dynamic risk feature time axis; Step D23: constructing a pre-risk dynamic analysis model based on the dynamic risk feature timeline to generate an initial risk prediction model; Step D24: Based on the transfer learning framework, the credit assessment model parameters in the source domain are transferred to the target domain, and the domain adaptation network is used to align the feature distribution; Step D25: Optimize the risk prediction model through adversarial training strategy and generate a deep neural network model, and then use the Bayesian optimization algorithm to automatically tune the hyperparameters of the deep neural network model to obtain the risk prediction optimization model.

4. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that: Step D3 includes the following steps: Step D31: Input the customer's complete feature matrix into the risk prediction optimization model to obtain the risk probability distribution at each time point; Step D32: Decompose the time-frequency characteristics of the risk probability distribution using a wavelet transform algorithm to extract the main frequency components of the risk fluctuations, and then construct a risk characteristic fluctuation map based on the main frequency components of the risk fluctuations; Step D33: Using an independent component analysis algorithm to perform multi-dimensional risk indicator decoupling analysis on the risk characteristic fluctuation map, thereby isolating key risk influencing factors, which include income fluctuation factors, debt expansion factors, and behavioral abnormality factors; Step D34: Construct a risk transmission network based on the system dynamics model to simulate the risk transmission paths of key risk factors in the capital chain, credit chain, and supply chain; Step D35: Perform sensitivity analysis through the risk transmission path to determine the warning threshold, and then generate a dynamic risk warning signal based on the warning threshold.

5. A consumer finance intelligent assisted credit granting method according to claim 4, characterized in that: Step D4 includes the following steps: Step D41: Adjust the branch weights of the credit decision tree based on the dynamic risk warning signal to build an adaptive decision model; Step D42: Stress-test the customer's repayment ability through the adaptive decision-making model to generate credit limit ranges under different economic scenarios; Step D43: Using fuzzy mathematics methods to calculate the membership of credit limit intervals under different economic scenarios, and determine the gradient credit limit interval; Step D44: Construct a dynamic adjustment matrix for decision weights, perform nonlinear mapping between dynamic risk warning signals and gradient credit limit intervals, and balance the risk-return ratio through a multi-objective optimization algorithm to generate an intelligent credit decision report.

6. A consumer finance intelligent assisted credit granting method according to claim 5, characterized in that: Multi-objective optimization algorithms include: Define the optimization objective function set, optimize the objective function set, generate a set of candidate solutions, sort the candidate solutions, identify the optimal compromise solution, apply robust optimization theory to perform sensitivity analysis on the optimal compromise solution, and generate an intelligent credit decision report based on the results of the sensitivity analysis.

7. A consumer finance intelligent assisted credit granting system, applied to a consumer finance intelligent assisted credit granting method according to any one of claims 1 to 6, characterized in that: include: The data acquisition and analysis module is used to obtain basic customer data, financial transaction flow data, and third-party credit data through multi-source heterogeneous data acquisition interfaces, construct an initial customer feature set, perform null-value feature vector mapping analysis on the initial customer feature set, generate a null-value feature distribution heat map, and perform multi-dimensional data intelligent completion based on the null-value feature distribution heat map to generate a complete customer feature matrix; The risk prediction optimization module is used to reorganize the time-varying feature sequence of the customer's complete feature matrix, construct a dynamic risk feature timeline, build a pre-risk dynamic analysis model based on the dynamic risk feature timeline, generate an initial risk prediction model, and migrate the model parameters of the initial risk prediction model across scenarios to obtain a risk prediction optimization model; The early warning module is used to perform real-time risk scanning of the customer's complete feature matrix based on the risk prediction optimization model, generate a risk feature fluctuation map, perform multi-dimensional risk indicator decoupling analysis on the risk feature fluctuation map, extract key risk influencing factors, simulate the risk transmission path based on the key risk influencing factors, and generate dynamic risk early warning signals; The report generation module is used to reconstruct the credit decision tree based on dynamic risk warning signals, generate an adaptive credit decision model, calculate the credit limit elasticity of the customer's complete feature matrix through the adaptive credit decision model, generate a gradient credit limit range, fuse the gradient credit limit range with the dynamic risk warning signal for decision weight, and output an intelligent credit decision report.

8. A computer-readable storage medium, characterized in that: Instructions are stored, and when the instructions are run on a computer, the computer is caused to execute a consumer finance intelligent assisted credit granting method as claimed in any one of claims 1 to 6.

Images