Unhealthy asset management method based on adaptive multi-modal causal reasoning network

Abstract

The invention discloses a non-performing asset management method based on an adaptive multi-modal causal inference network. The method comprises the steps of data acquisition and integration, data preprocessing, data labeling and classification, causal relationship adaptive identification module proposing, model training and optimization, and prediction and management. The causal relationship self-adaptive identification module is used for constructing an efficient and dynamically updated causal graph through a Bayesian network learning method and a method of integrating prior knowledge and data driving; wherein the causal relationship adaptive identification module comprises hierarchical Bayesian prior and knowledge graph fusion, a dynamic adjustment mechanism, hierarchical variational distribution and adaptive regular interaction. According to the method, the accuracy of the bad asset prediction can be remarkably improved, and the defects of a traditional statistical model and a machine learning algorithm in complex relation capture are overcome. The model interpretability is enhanced, and a dynamic adjustment mechanism ensures that the model can respond to market changes in real time.

Description

Technical Field

[0001] The present invention relates to the financial field, and particularly to a non-performing asset management method based on an adaptive multi-modal causal inference network. Background Art

[0002] With the continuous improvement of the complexity and dynamics of the financial market, financial institutions are facing increasingly severe challenges in non-performing asset management. Currently, in the prediction and management of non-performing assets, financial institutions mainly adopt the following types of methods: Statistical models: such as Logistic regression, linear discriminant analysis, etc. Although they perform well in modeling simple linear relationships, it is difficult to capture complex non-linear relationships and high-dimensional data features. Machine learning algorithms: such as decision trees, random forests, support vector machines, etc. These methods have improved in dealing with high-dimensional data and non-linear relationships, but usually ignore the causal relationships between data, resulting in insufficient interpretability of the models. Deep learning models: such as multi-layer perceptrons, convolutional neural networks, recurrent neural networks, etc. Although they perform excellently on large-scale data, their "black box" characteristics make it difficult to achieve causal inference, restricting the application of the models in the financial field. Traditional non-performing asset management methods mainly rely on statistical analysis and rule-based systems. When facing multi-source heterogeneous data and complex causal relationships, these methods often show problems of insufficient prediction accuracy and poor model interpretability.

[0003] Therefore, there is an urgent need for an advanced algorithm that can deeply mine the internal causal relationships of data and combine multi-modal data for accurate prediction to improve the efficiency and effectiveness of non-performing asset management. For this reason, the present invention proposes an innovative algorithm framework based on an adaptive multi-modal causal inference network, aiming to improve the accuracy of non-performing asset prediction and the robustness of the system by dynamically identifying and modeling the causal relationships in financial data. This solution combines causal inference with multi-modal data fusion to achieve accurate prediction and efficient management of non-performing assets in a complex financial environment. Summary of the Invention

[0004] The purpose of the present invention is to provide a non-performing asset management method based on an adaptive multi-modal causal inference network to solve the problems existing in the background art.

[0005] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0006] A non-performing asset management method based on an adaptive multi-modal causal inference network, the method comprising:

[0007] S1. Data collection and integration: Collect multi-source data, where the multi-source data includes structured data, unstructured data, and time-series data;

[0008] S2. Data preprocessing: Data cleaning and data standardization. Data cleaning includes handling missing values and outliers.

[0009] S3. Data annotation and classification: Define labels and the annotation process. In non-performing asset prediction, manually annotate data as labels for model training.

[0010] S4. Propose a Causal Relationship Adaptive Recognition Module (CRARM); this module, through Bayesian network learning methods, innovatively integrates prior knowledge and data-driven methods to construct an efficient and dynamically updated causal graph; among them, the causal relationship adaptive recognition module includes the integration of hierarchical Bayesian priors and knowledge graphs, a dynamic adjustment mechanism (DAM), a hierarchical variational distribution (HVD), and adaptive regular interactions.

[0011] S5. Model training and optimization: Initialize the model and the training process. Among them, initialize the model: Based on the causal graph and prior probabilities, set the initial model parameters; the training process includes: variational inference, Bayesian optimization, and alternating optimization; variational inference: Optimize the model to better explain the causal relationships in the data; Bayesian optimization: Adjust the model hyperparameters to improve the prediction performance; alternating optimization: Continuously optimize the causal graph and model parameters during the training process to ensure the robustness and accuracy of the model.

[0012] S6. Prediction and management: Use the trained model to predict the non-performing asset risks of new or existing customers. According to the prediction results, formulate corresponding risk management strategies, such as adjusting loan conditions, strengthening collection measures, etc.; Monitor market and customer data in real time, and dynamically update the model and the causal graph to ensure the continuous effectiveness of the prediction.

[0013] As a further technical solution of the present invention, the method further includes system maintenance and optimization and model deployment and real-time prediction. Among them, the system maintenance and optimization includes:

[0014] Regular evaluation: Regularly evaluate the performance of the model to ensure its applicability in different market environments.

[0015] Feedback mechanism: Continuously optimize the algorithm and model structure according to the actual management effects and new data feedback.

[0016] The model deployment and real-time prediction includes:

[0017] Goal: Deploy the trained model to the production environment to achieve real-time non-performing asset prediction and monitoring.

[0018] Export the trained model into a deployable format;

[0019] Build a real-time data flow pipeline to ensure that new data can be input into the model for prediction in a timely manner. Integrate the model into the risk management system of financial institutions to achieve automated loan default prediction. Configure a real-time monitoring and alarm mechanism to respond to anomalies in the prediction results in a timely manner;

[0020] Continuous monitoring and maintenance:

[0021] Monitor the prediction performance of the model in real time, collect feedback data for subsequent optimization, and regularly update the causal graph and model parameters to ensure that the system can adapt to market changes.

[0022] As a further technical solution of the present invention, the processing of missing values in step S2 includes the following steps:

[0023] For samples or features with a missing value ratio exceeding a set threshold (such as 30%), directly delete them;

[0024] ;

[0025] Among them, : The filled data, Mean(X): The mean of variable X, X: The original data;

[0026] For those below the threshold, use the Multiple Imputation method to predict the missing values based on the values of other variables;

[0027] ;

[0028] Among them, and are the mean and standard deviation of variable X respectively, and Z is a standard normal distribution random variable.

[0029] As a further technical solution of the present invention, the outlier processing in step S2 includes the following steps:

[0030] In the process of non-performing asset prediction, outliers may significantly affect the prediction accuracy of the model. Therefore, the CRARM module adopts a comprehensive outlier detection and processing method. First, calculate the Z-score of each data point, and the formula is:

[0031] ;

[0032] Among them, Z: Standard score, μ: Mean of variable X, : Standard deviation of variable X. When , mark this data point as an outlier. In addition, use the interquartile range (IQR), that is:

[0033] IQR = Q3 - Q1;

[0034] where Q1: the first quartile, Q3: the third quartile;

[0035] Further detect outliers. When a data point is lower than or higher than it is also regarded as an outlier. For the detected outliers, the CRARM module adopts a capping strategy to process them, that is, the values exceeding the upper and lower limits are respectively replaced with

[0036] .

[0037] As a further technical solution of the present invention, the data annotation and classification in step S3 include the following steps:

[0038] Label definition:

[0039] Normal: The loan customer repays the loan on time without any default records;

[0040] Overdue: The loan customer has an overdue repayment record;

[0041] Default: The loan customer fails to repay the loan on schedule and officially enters the default state;

[0042] ;

[0043] Annotation method:

[0044] Rule benchmark: Make a preliminary annotation according to the preset financial index threshold;

[0045] Manual review: Financial analysts review and correct the preliminary annotation results to ensure the accuracy of the annotation;

[0046] ;

[0047] Label verification:

[0048] Cross-validation: Adopt the cross-validation method to ensure the consistency and reliability of the annotated data;

[0049] Consistency check: Conduct a consistency check on multi-label data to avoid annotation conflicts;

[0050] Stability score = frequency of consistent annotation / total cumulative detections;

[0051] If the stability score is lower than the preset threshold (such as 0.8), the corresponding causal edge is removed.

[0052] As a further technical solution of the present invention, the fusion of hierarchical Bayesian prior and knowledge graph in step S4 includes the following steps:

[0053] First, using the knowledge graph in the financial field, systematically represent the domain expert knowledge and literature research results as a graph structure. The nodes V in the knowledge graph represent financial variables, the edges E represent known causal relationships or influence paths, and G represents the knowledge graph:

[0054] ;

[0055] Based on the causal relationships in the knowledge graph, set the high-level prior probability distribution. For the existing causal edges , assign a higher prior probability ; for the non-existing edges, assign a lower probability , where ;

[0056] ;

[0057] where P(e) is the prior probability of the causal edge e, E: the set of known causal edges in the knowledge spectrum graph, is the high prior probability of the known causal edge, is the low prior probability of the unknown causal edge. On the basis of the high-level prior, CRARM introduces data-driven adjustment of the causal edge probability. Through the Bayesian update formula, integrate the statistical information in the data into the prior distribution to form a more accurate posterior distribution;

[0058] ;

[0059] is the posterior probability of the causal edge e under the observed data D;

[0060] is the likelihood of the data D under the condition that the causal edge e exists;

[0061] P(e) is the prior probability of the causal edge e;

[0062] P(D) is the marginal likelihood of the data D, used for normalization.

[0063] Through this hierarchical prior setting and adjustment, the CRARM module can effectively combine prior knowledge and data-driven information to improve the accuracy and adaptability of causal graph construction.

[0064] As a further technical solution of the present invention, the dynamic adjustment mechanism in step S4 includes the following steps:

[0065] Real-time data analysis and knowledge graph update: Combine the results of real-time data analysis with the latest insights of domain experts to dynamically update the causal relationships in the knowledge graph. This process includes adding new causal edges and removing edges that are no longer applicable;

[0066] Reset high-level prior probabilities: According to the updated knowledge graph, reset the high-level prior probability P(e) to ensure that the prior information always reflects the latest domain knowledge and market conditions;

[0067] Bayesian optimization: Use the Bayesian optimization algorithm to automatically adjust the high-level prior parameters and to optimally integrate prior knowledge and data-driven information.

[0068] As a further technical solution of the present invention, the hierarchical variational distribution in step S4 includes the following steps:

[0069] Define the variational distribution: Define the variational distribution q(e) of the causal edge e, where e represents the existence or non-existence of the causal edge. Assuming that each causal edge is independent, the variational distribution can be expressed as the product of the independent distributions of each edge:

[0070] ;

[0071] where q(G) is the variational distribution of the causal graph G, and q(e) is the variational distribution of the causal edge e, representing the probability of its existence;

[0072] Parameterize the existence probability: For each causal edge e, use a deep neural network (such as a multi-layer perceptron) to parameterize its existence probability; specifically, define the variational distribution of each causal edge as a Bernoulli distribution, and its parameter is determined by the Sigmoid function value output by the neural network:

[0073] ;

[0074] where q(e): the variational distribution of the causal edge e, and are the weight matrix and bias of the causal edge e respectively, is the Sigmoid activation function to ensure that the output probability is between (0,1);

[0075] CRARM introduces a dynamic weight adjustment mechanism to dynamically adjust the weights of the prior and likelihood according to the real-time feedback of the data, ensuring that the model can adapt to the new market conditions in a timely manner; Introduce a dynamic weight parameter whose value changes with time t:

[0076] ;

[0077] where, is a time-varying weight parameter used to balance the influence of the prior and the likelihood. γ controls the speed of weight adjustment and determines the rate of gradual transition from prior dominance to data-driven, and when t increases, it gradually increases from close to 0 to close to 1, indicating that the model increasingly relies on data-driven information.

[0078] As a further technical solution of the present invention, the adaptive regular interaction in step S4 includes the following steps:

[0079] Introduce a regularization weight matrix into the variational lower bound , whose elements are dynamically adjusted according to the importance of the causal edge e:

[0080] ;

[0081] is the Variational Lower Bound, which is used to optimize causal relationship recognition, is the expected likelihood, which measures the reconstruction ability of data under the current causal graph, is the KL divergence, which measures the difference between the variational distribution q(e) and the prior distribution P(e), is the regularization weight of the causal edge e, which is calculated through the importance score S(e) of the causal edge:

[0082] ;

[0083] is the basic regularization strength, is the importance score of the causal edge e (such as the weight based on the knowledge graph or the data-driven edge weight);

[0084] At the same time, considering that traditional causal learning methods usually assume that causal relationships are one-way and fixed. CRARM introduces interactive causal inference, allowing interactive effects between causal edges. For example, the presence of certain causal edges may enhance or inhibit the influence of other edges. This is achieved by introducing the interactive parameter of the causal edge:

[0085] ;

[0086] where is further modeled as:

[0087] ;

[0088] where is given the parent node Under the causal graph G, the conditional probability of the variable is the data variable corresponding to the parent node j in the causal graph, is the weight from the parent node j to the node i, is the interaction parameter between causal edges, indicating the influence of causal edge k on causal edge j, is related to the interaction parameter associated data variable, is the Sigmoid function, which is used to map the result of the linear combination to the range (0, 1). In this way, CRARM not only captures direct causal relationships but also can identify more complex causal patterns through interactions.

[0089] As a further technical solution of the present invention, step S5 includes the following steps:

[0090] The joint learning objective of the causal relationship adaptive recognition module is expressed as:

[0091] ;

[0092] where is the variational lower bound, which is used as the optimization objective of joint learning, is the dynamic weight parameter, which balances the weights of the likelihood term and the prior term, is the expected likelihood, which measures the reconstruction ability of the data under the current causal graph, is the expected prior, which reflects the prior causal relationship in the knowledge graph, is the KL divergence of the causal edge e, which measures the difference between the variational distribution and the prior distribution, is the regularization weight of the causal edge e, which adjusts the regularization intensity based on its importance score;

[0093] The optimization process is as follows:

[0094] Initialization: Based on the knowledge graph and the prior probability matrix P(G), initialize the causal graph G;

[0095] Variational inference: In each iteration step, first optimize the variational lower bound through variational inference ;

[0096] Bayesian optimization: Secondly, use Bayesian optimization to adjust the model hyperparameter θ;

[0097] Dynamic update: According to the optimization results, dynamically update the causal graph G and the weight matrix W to ensure the real-time and accuracy of causal relationships.

[0098] The beneficial effects achieved by the present invention:

[0099] Improving prediction accuracy: By deeply mining and modeling causal relationships and combining multi-modal data, the present invention can significantly improve the accuracy of non-performing asset prediction, overcoming the deficiencies of traditional statistical models and machine learning algorithms in capturing complex relationships.

[0100] Enhancing model interpretability: The introduction of causal reasoning enables the model to not only have high prediction performance but also good interpretability, providing a more reliable basis for financial decision-making and solving the "black box" problem of deep learning models.

[0101] Real-time dynamic adaptation: The dynamic adjustment mechanism ensures that the model can respond to market changes in real time, maintaining the timeliness and accuracy of the causal graph and model parameters, adapting to the rapid changes in the financial market, and enhancing the robustness of the system. Brief Description of the Drawings

[0102] Figure 1 It is a flowchart of a non-performing asset management method based on an adaptive multi-modal causal inference network. Detailed Embodiment

[0103] The technical solutions of the present invention will be described in detail below with reference to specific drawings.

[0104] Please refer to Figure 1 , the embodiment of the present invention provides a non-performing asset management method based on an adaptive multi-modal causal inference network, and the method includes:

[0105] S1. Data collection and integration: Collect multi-source data, where the multi-source data includes structured data (such as customer information, credit scores), unstructured data (such as financial news, industry reviews), and time-series data (such as transaction records, market prices). Table 1 below details the categories, sources, and specific examples of various types of data;

[0106] Table 1 - Specific Examples of Multi-Source Data

[0107]

[0108] S2. Data preprocessing: Data cleaning and data standardization. Data cleaning includes handling missing values and outliers. Handling missing values: Delete samples or features with a high proportion of missing values (such as exceeding a set threshold of 30%), and use imputation methods to fill the remaining missing values. Outlier handling: Identify and correct outliers in the data to ensure the rationality of the data distribution. Data standardization refers to unifying the scales of different types of data for subsequent analysis and modeling;

[0109] Handling missing values includes the following steps:

[0110] For samples or features with a missing value ratio exceeding a set threshold (such as 30%), directly delete them;

[0111] ;

[0112] Among them, : The filled data, mean (X): The mean of variable X, X: The original data;

[0113] For those below the threshold, the Multiple Imputation method is adopted to predict the missing values based on the values of other variables;

[0114] ;

[0115] Among them, and are the mean and standard deviation of variable X respectively, and Z is a standard normal distribution random variable.

[0116] Outlier handling includes the following steps:

[0117] During the non-performing asset prediction process, outliers may significantly affect the prediction accuracy of the model. Therefore, the CRARM module adopts a comprehensive outlier detection and handling method. First, by calculating the Z-score of each data point, the formula is:

[0118] ;

[0119] Among them, Z: Standard score, μ: Mean of variable X, : Standard deviation of variable X, when , mark this data point as an outlier. In addition, use the interquartile range (IQR), that is:

[0120] IQR = Q3 - Q1;

[0121] Among them, Q1: The first quartile (25% quantile), Q3: The third quartile (75% quantile);

[0122] Further detect outliers. When the data point is below or above , it is also regarded as an outlier. For the detected outliers, the CRARM module adopts a capping strategy for processing, that is, replace the values exceeding the upper and lower limits with

[0123] ;

[0124] This method not only effectively eliminates the negative impact of outliers on model training but also maintains the overall distribution characteristics of the data, thus ensuring the efficient operation of subsequent causal relationship identification and prediction models. For example, when processing customer loan amount data, if a customer's loan amount is 10 million yuan and the upper limit determined by the IQR method is 6 million yuan, then this outlier will be truncated to 6 million yuan to prevent it from causing excessive bias to the model.

[0125] S3. Data annotation and classification: Define labels and the annotation process. In non-performing asset prediction, manually annotate data as labels for model training. The accuracy of data annotation directly affects the performance of the model;

[0126] The data annotation and classification in step S3 include the following steps:

[0127] Label definition:

[0128] Normal: The loan customer repays on time without default records;

[0129] Overdue: The loan customer has an overdue repayment record;

[0130] Default: The loan customer fails to repay on schedule and officially enters the default state;

[0131] ;

[0132] Annotation method:

[0133] Rule-based benchmark: Conduct preliminary annotation based on preset financial indicator thresholds; for example, if the overdue days exceed 30 days, it is annotated as "overdue", and if it exceeds 90 days, it is annotated as "default";

[0134] Manual review: Financial analysts review and correct the preliminary annotation results to ensure the accuracy of annotation;

[0135] ;

[0136] Label verification:

[0137] Cross-validation: Adopt the cross-validation method to ensure the consistency and reliability of the annotated data;

[0138] Consistency check: Conduct a consistency check on multi-label data to avoid annotation conflicts;

[0139] Stability score = frequency of consistent annotation / total cumulative detections;

[0140] If the stability score is lower than the preset threshold (such as 0.8), then remove the corresponding causal edge.

[0141] Data annotation example:

[0142] In one embodiment, taking the loan customers of a certain financial institution as an example, the collected data includes loan amount, term, interest rate, repayment method, credit score, payment history, etc. At a certain point in time, it is observed that the loan customer has an overdue repayment record and the credit score has decreased, and this data point is manually labeled as the "overdue" status.

[0143] In the prediction of non-performing financial assets, accurately identifying the causal relationships between variables is crucial for improving the prediction accuracy and interpretability of the model. Traditional methods mostly rely on correlation analysis. However, causal relationships reveal the direct influence paths between variables, which helps to identify key driving factors, reduce model bias, and improve the accuracy of prediction and the reliability of decision-making. For example, the impact of interest rate changes on the loan default rate is not only reflected in the correlation, but also an important manifestation of the causal relationship. To achieve more efficient, dynamic, and accurate causal relationship identification, the present invention proposes a causal relationship adaptive recognition module.

[0144] S4. Propose a Causal Relationship Adaptive Recognition Module (CRARM); this module innovatively integrates prior knowledge and data-driven methods through Bayesian network learning to construct an efficient and dynamically updated causal graph; wherein, the causal relationship adaptive recognition module includes hierarchical Bayesian prior and knowledge graph fusion, a dynamic adjustment mechanism (DAM), a hierarchical variational distribution (HVD), and adaptive regular interaction.

[0145] The hierarchical Bayesian prior and knowledge graph fusion in step S4 includes the following steps:

[0146] First, using the knowledge graph in the financial field, systematically represent the domain expert knowledge and literature research results as a graph structure. The nodes V in the knowledge graph represent financial variables, the edges E represent known causal relationships or influence paths, and G represents the knowledge graph.

[0147] ;

[0148] Based on the causal relationships in the knowledge graph, set the high-level prior probability distribution. For the existing causal edges , assign a higher prior probability ; for the non-existing edges, assign a lower probability , where ;

[0149] ;

[0150] where \(P(e)\) is the prior probability of causal edge \(e\), and \(E\) is the set of known causal edges in the knowledge spectral graph. is the high prior probability of known causal edges (e.g., 0.9), and \(P'(e)\) is the low prior probability of unknown causal edges (e.g., 0.1). Based on the high-level prior, CRARM introduces data-driven adjustment of causal edge probabilities. Through the Bayesian update formula, the statistical information in the data is incorporated into the prior distribution to form a more accurate posterior distribution.

[0151] ;

[0152] \(P(e|D)\) is the posterior probability of causal edge \(e\) under the observed data \(D\);

[0153] \(P(D|e)\) is the likelihood of data \(D\) given the existence of causal edge \(e\);

[0154] \(P(e)\) is the prior probability of causal edge \(e\);

[0155] \(P(D)\) is the marginal likelihood of data \(D\) for normalization.

[0156] Through this hierarchical prior setting and adjustment, the CRARM module can effectively combine prior knowledge and data-driven information, improving the accuracy and adaptability of causal graph construction.

[0157] The dynamic adjustment mechanism in step S4 includes the following steps:

[0158] Real-time data analysis and knowledge graph update: Combining the results of real-time data analysis and the latest insights of domain experts, dynamically update the causal relationships in the knowledge graph. This process includes adding new causal edges and removing edges that are no longer applicable;

[0159] Reset the high-level prior probability: According to the updated knowledge graph, reset the high-level prior probability \(P(e)\) to ensure that the prior information always reflects the latest domain knowledge and market conditions;

[0160] Bayesian optimization: Use the Bayesian optimization algorithm to automatically adjust the high-level prior parameters and to fuse prior knowledge and data-driven information in the best way.

[0161] The hierarchical variational distribution in step S4 includes the following steps:

[0162] Define the variational distribution: Define the variational distribution \(q(e)\) of causal edge \(e\), where \(e\) represents the existence or non-existence of the causal edge. Assuming that each causal edge is independent, the variational distribution can be expressed as the product of independent distributions of each edge:

[0163] ;

[0164] where \(q(G)\) is the variational distribution of the causal graph \(G\), and \(q(e)\) is the variational distribution of the causal edge \(e\), representing the probability of its existence;

[0165] Parameterize the existence probability: For each causal edge \(e\), use a deep neural network (such as a multi-layer perceptron) to parameterize its existence probability; specifically, define the variational distribution of each causal edge as a Bernoulli distribution, and its parameter is determined by the Sigmoid function value output by the neural network:

[0166] ;

[0167] where \(q(e)\): the variational distribution of the causal edge \(e\), and are the weight matrix and bias of the causal edge \(e\) respectively, is the Sigmoid activation function, ensuring that the output probability is between (0, 1). Through HVD, CRARM can effectively capture the potential patterns of causal relationships and achieve accurate causal edge identification in high-dimensional data;

[0168] In the traditional Bayesian framework, the weights of the prior and likelihood functions are usually fixed, which may lead to the model being less flexible in responding to new information in the rapidly changing environment of the financial market. CRARM introduces a dynamic weight adjustment mechanism to dynamically adjust the weights of the prior and likelihood according to the real-time feedback of the data, ensuring that the model can adapt to new market conditions in a timely manner; introduce a dynamic weight parameter , whose value changes with time :

[0169] ;

[0170] where is the time-varying weight parameter used to balance the influence of the prior and likelihood, and \(\gamma\) controls the speed of weight adjustment, determining the rate of gradual transition from prior dominance to data-driven, and when \(t\) increases, gradually increases from close to 0 to close to 1, indicating that the model gradually relies more on data-driven information.

[0171] The adaptive regularization interaction in step S4 includes the following steps:

[0172] To prevent overfitting of the causal graph in high-dimensional data, the present invention introduces an adaptive regularization interaction mechanism in CRARM. This mechanism adaptively adjusts the regularization strength according to the importance of the causal edges, giving a lower regularization strength to the key causal edges and imposing a higher regularization on the unimportant or uncertain causal edges. Specifically, a regularization weight matrix is introduced in the variational lower bound , whose elements are dynamically adjusted according to the importance of the causal edge e:

[0173] ;

[0174] is the Variational Lower Bound, which is used to optimize causal relationship recognition. is the expected likelihood, which measures the reconstruction ability of the data under the current causal graph. is the KL divergence, which measures the difference between the variational distribution q(e) and the prior distribution P(e). is the regularization weight of the causal edge e, which is calculated through the importance score S(e) of the causal edge:

[0175] ;

[0176] is the basic regularization strength, and S(e) is the importance score of the causal edge e (such as the weight based on the knowledge graph or the data-driven edge weight);

[0177] At the same time, considering that traditional causal learning methods usually assume that causal relationships are one-way and fixed. CRARM introduces interactive causal inference, allowing interactive effects between causal edges. For example, the existence of certain causal edges may enhance or inhibit the influence of other edges. This is achieved by introducing the interactive parameter of the causal edge:

[0178] ;

[0179] Among them, is further modeled as:

[0180] ;

[0181] Among them is the conditional probability of the variable given the parent node and the causal graph G. is the data variable corresponding to the parent node j in the causal graph. is the weight from the parent node j to the node i. is the interactive parameter between causal edges, indicating the influence of the causal edge k on the causal edge j. is the interactive parameter associated data variable. is the Sigmoid function, which is used to map the result of the linear combination to the range (0, 1). In this way, CRARM not only captures the direct causal relationship but also can identify more complex causal patterns through interactions.

[0182] The causal relationship identification and modeling of the present invention include integrating prior knowledge: using the knowledge graph in the financial field to initially construct a causal relationship network. Data-driven adjustment: Dynamically identify and update the causal relationship through an algorithm to ensure that the model reflects the latest data and market changes. Dynamic update: Continuously optimize the causal graph according to real-time data and expert opinions to maintain the accuracy and adaptability of the model.

[0183] S5. Model training and optimization: Initialize the model and the training process. Among them, initialize the model: Based on the causal graph and prior probability, set the initial model parameters; the training process includes: variational inference, Bayesian optimization, and alternating optimization; Variational inference: Optimize the model to better explain the causal relationship in the data; Bayesian optimization: Adjust the model hyperparameters to improve the prediction performance; Alternating optimization: Continuously optimize the causal graph and model parameters during the training process to ensure the robustness and accuracy of the model;

[0184] Step S5 includes the following steps:

[0185] The joint learning objective of the causal relationship adaptive identification module is expressed as:

[0186] ;

[0187] Where is the variational lower bound, which is used as the optimization objective of joint learning, is the dynamic weight parameter, which balances the weights of the likelihood term and the prior term, is the expected likelihood, which measures the reconstruction ability of the data under the current causal graph, is the expected prior, which reflects the prior causal relationship in the knowledge graph, is the KL divergence of the causal edge e, which measures the difference between the variational distribution and the prior distribution, is the regularization weight of the causal edge e, which adjusts the regularization strength based on its importance score;

[0188] The optimization process is as follows:

[0189] Initialization: Based on the knowledge graph and the prior probability matrix P(G), initialize the causal graph G;

[0190] Variational inference: In each iteration step, first optimize the variational lower bound ;

[0191] Bayesian optimization: Secondly, use Bayesian optimization to adjust the model hyperparameters (such as and );

[0192] Dynamic update: According to the optimization results, dynamically update the causal graph G and the weight matrix W to ensure the real-time and accuracy of causal relationships.

[0193] S6, Prediction and management: Use the trained model to predict the non-performing asset risks of new or existing customers. According to the prediction results, formulate corresponding risk management strategies, such as adjusting loan conditions, strengthening collection measures, etc.; Monitor the market and customer data in real time, and dynamically update the model and causal graph to ensure the continuous effectiveness of the prediction.

[0194] The method also includes system maintenance and optimization and model deployment and real-time prediction. Among them, the system maintenance and optimization includes:

[0195] Regular evaluation: Regularly evaluate the performance of the model to ensure its applicability in different market environments;

[0196] Feedback mechanism: Continuously optimize the algorithm and model structure according to the actual management effect and new data feedback;

[0197] The model deployment and real-time prediction includes:

[0198] Objective: Deploy the trained model to the production environment to achieve real-time non-performing asset prediction and monitoring;

[0199] Export the trained model into a deployable format (such as TensorFlow SavedModel, ONNX, etc.);

[0200] Build a real-time data flow pipeline to ensure that new data can be input into the model for prediction in a timely manner. Integrate the model into the risk management system of financial institutions to achieve automated loan default prediction. Configure a real-time monitoring and alarm mechanism to respond to abnormal situations in the prediction results in a timely manner;

[0201] Continuous monitoring and maintenance:

[0202] Monitor the prediction performance of the model in real time, collect feedback data for subsequent optimization, and regularly update the causal graph and model parameters to ensure that the system can adapt to market changes.

[0203] Example:

[0204] The deployed model automatically receives the latest customer data and market metrics every day and outputs the loan default probability in real time.

[0205] When the model detects that the default probability of a certain customer has increased abnormally, the system automatically triggers a risk alarm to remind the relevant departments to take measures.

[0206] The hierarchical Bayesian prior of the present invention is integrated with a knowledge graph, innovatively combining the knowledge graph in the financial field with the hierarchical Bayesian prior, systematically integrating domain expert knowledge and data-driven information for constructing and optimizing a causal graph. This method significantly improves the accuracy and adaptability of causal relationship recognition.

[0207] Hierarchical variational distribution (HVD): The existence probability of each causal edge is parameterized using a deep neural network. By decomposing the joint distribution of causal edges into independent distributions, efficient recognition of causal relationships in high-dimensional data is achieved. HVD effectively captures the complex relationships and potential patterns of causal edges, improving the precision of causal relationship recognition.

[0208] Interactive causal inference mechanism: Allows for interactive effects between causal edges. By introducing interaction parameters for causal edges, more complex causal patterns are identified and modeled. This not only captures direct causal relationships but also reveals deeper causal structures through interactions, enhancing the flexibility and accuracy of the model.

[0209] It should be noted that in this article, the term "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device including a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device including that element.

[0210] The above are only the preferred embodiments of the present invention and do not limit the scope of the invention of the present invention. Any equivalent structural or equivalent process transformation made using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included within the scope of the invention protection of the present invention.

Claims

1. A non-performing asset management method based on an adaptive multimodal causal inference network, characterized in that: The method comprises: S1. Data collection and integration: Collect multi-source data, including structured data, unstructured data and time series data; S2. Data preprocessing: data cleaning and data standardization. Data cleaning includes processing missing values ​​and outliers. S3, Data labeling and classification: define labels and labeling process. In the prediction of non-performing assets, manually labeled data is used as labels for model training; S4. Propose a causal relationship adaptive identification module; through the Bayesian network learning method, integrate prior knowledge and data-driven methods to build an efficient and dynamically updated causal graph; wherein the causal relationship adaptive identification module includes hierarchical Bayesian prior and knowledge graph fusion, dynamic adjustment mechanism, hierarchical variational distribution, and adaptive regular interaction; S5, model training and optimization: initializing the model and the training process, wherein initializing the model: setting the initial model parameters based on the causal graph and prior probability; the training process includes: variational inference, Bayesian optimization and alternating optimization; S6. Prediction and Management: Use the trained model to predict the client’s non-performing asset risk and formulate corresponding risk management strategies based on the prediction results; monitor the market and client data in real time, dynamically update the model and causal map, and ensure the continued effectiveness of the prediction.

2. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The method also includes system maintenance and optimization and model deployment and real-time prediction, wherein the system maintenance and optimization includes: Regular evaluation: Regularly evaluate the performance of the model to ensure its applicability in different market environments; Feedback mechanism: Continuously optimize algorithms and model structures based on actual management results and new data feedback; The model deployment and real-time prediction include: Objective: Deploy the trained model to the production environment to achieve real-time non-performing asset prediction and monitoring; Export the trained model into a deployable format; Build a real-time data flow pipeline to ensure that new data is input into the model for prediction in a timely manner, integrate the model into the risk management system of the financial institution to realize automated loan default prediction, configure real-time monitoring and alarm mechanisms, and respond to abnormal situations in the prediction results in a timely manner; Ongoing monitoring and maintenance: Monitor the model's predictive performance in real time, collect feedback data for subsequent optimization, and regularly update causal graphs and model parameters to ensure that the system adapts to market changes.

3. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: Processing missing values ​​in step S2 includes the following steps: For samples or features whose missing value ratio exceeds the set threshold, delete them directly; ； in, : filled data, mean(X): mean of variable X, X: original data; If the value is below the threshold, multiple imputation method is used to predict the missing value based on the value of other variables; ； in, and are the mean and standard deviation of variable X, respectively, and Z is a standard normally distributed random variable.

4. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The outlier processing in step S2 includes the following steps: By calculating the Z-score of each data point, the formula is: ； Among them, Z: standard score, μ: mean of variable X, : The standard deviation of variable X, when When , the data point is marked as abnormal. In addition, the interquartile range is used, that is: IQR=Q3-Q1; Among them, Q1: first quartile, Q3: third quartile; Further outlier detection, when the data point is below or higher , it is also regarded as an outlier. For the detected outliers, a truncation strategy is used to process them, that is, the values ​​that exceed the upper and lower limits are replaced by: 。 5. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The data labeling and classification in step S3 includes the following steps: Tag definition: Normal: The loan customer repays the loan on time and has no record of default; Overdue: The loan customer has a record of overdue repayment; Default: The loan customer fails to repay the loan on time and officially enters the default state; ； Marking method: Rule benchmark: preliminary marking based on preset financial indicator thresholds; Manual review: Financial analysts review and revise the preliminary annotation results to ensure the accuracy of the annotations; ； Tag Verification: Cross-validation: Use cross-validation method to ensure the consistency and reliability of labeled data; Consistency check: Perform consistency check on multi-label data to avoid labeling conflicts; Stability score = frequency of consistent labeling / total cumulative testing volume; If the stability score is lower than the preset threshold, the corresponding causal edge is removed.

6. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The fusion of hierarchical Bayesian prior and knowledge graph in step S4 includes the following steps: First, we use the knowledge graph in the financial field to systematically represent the domain expert knowledge and literature research results as a graph structure. The node V in the knowledge graph represents the financial variable, the edge E represents the known causal relationship or impact path, and G represents the knowledge graph: ； Based on the causal relationship in the knowledge graph, a high-level prior probability distribution is set for the existing causal edges. , giving a higher prior probability ; For non-existent edges, assign a lower probability ,in ; ； Where P(e) is the prior probability of the causal edge e, E: the set of known causal edges in the knowledge graph, is a high prior probability of a known causal edge, It is the low prior probability of unknown causal edges. On the basis of high-level priors, data-driven adjustment of causal edge probabilities is introduced. Through the Bayesian update formula, the statistical information in the data is integrated into the prior distribution to form a more accurate posterior distribution. ； is the posterior probability of the causal edge e under the observed data D; is the likelihood of data D given the existence of causal edge e; P(e) is the prior probability of the causal edge e; P(D) is the marginal likelihood of the data D, used for normalization.

7. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The dynamic adjustment mechanism in step S4 includes the following steps: Real-time data analysis and knowledge graph update: Combine real-time data analysis results with the latest insights from domain experts to dynamically update the causal relationships in the knowledge graph. This process includes adding new causal edges and removing edges that are no longer applicable. Reset high-level prior probabilities: Reset high-level prior probabilities P(e) based on the updated knowledge graph to ensure that the prior information always reflects the latest domain knowledge and market conditions; Bayesian Optimization: Automatically adjust high-level prior parameters using Bayesian optimization algorithms and , which optimally fuses prior knowledge with data-driven information.

8. The method for managing non-performing assets based on an adaptive multimodal causal inference network according to claim 1, characterized in that: The hierarchical variational distribution in step S4 includes the following steps: Define variational distribution: Define the variational distribution q(e) of the causal edge e, where e represents the existence or non-existence of the causal edge. Assuming that each causal edge is independent, the variational distribution can be expressed as the product of the independent distributions of each edge: ； Where q(G) is the variational distribution of the causal graph G, and q(e) is the variational distribution of the causal edge e, indicating its probability of existence; Parameterized existence probability: For each causal edge e, a deep neural network is used to parameterize its existence probability; the variational distribution of each causal edge is defined as a Bernoulli distribution, and its parameters are determined by the Sigmoid function value output by the neural network: ； Among them, q(e): the variational distribution of the causal edge e, and are the weight matrix and bias of the causal edge e, respectively. Sigmoid activation function ensures that the output probability is between (0,1); By introducing a dynamic weight adjustment mechanism, the weights of priors and likelihoods are dynamically adjusted according to real-time feedback from data, ensuring that the model adapts to new market conditions in a timely manner; Introducing a dynamic weight parameter , whose value changes with time t: ； in, is a weight parameter that changes over time and is used to balance the influence of priors and likelihoods. γ controls the speed of weight adjustment and determines The rate of gradual transition from prior-driven to data-driven, as t increases, It gradually increases from close to 0 to close to 1, indicating that the model gradually relies more on data-driven information.

9. The non-performing asset management method based on the adaptive multimodal causal reasoning network according to claim 1 is characterized in that: The adaptive regular interaction in step S4 includes the following steps: Introducing a regularized weight matrix in the variational lower bound , whose elements Dynamically adjust according to the importance of causal edge e: ； is a variational lower bound, used to optimize causal identification, is the expected likelihood, which measures the ability of data to be reconstructed under the current causal graph. is the KL divergence, which measures the difference between the variational distribution q(e) and the prior distribution P(e), is the regularized weight of the causal edge e, which is calculated by the importance score S(e) of the causal edge: ； is the basic regularization strength, Score the importance of causal edge e; Introduced interactive causal inference, allowing for interactive influences between causal edges and introducing interactive parameters for causal edges To achieve: ； in, Further modeling as: ； in At a given parent node and causal graph G, the variable The conditional probability of is the data variable corresponding to the parent node j in the causal graph, is the weight from parent node j to node i, is the interaction parameter between causal edges, indicating the influence of causal edge k on causal edge j. is the interaction parameter The associated data variables, It is the Sigmoid function, which is used to map the result of the linear combination to between (0,1).

10. The non-performing asset management method based on the adaptive multimodal causal inference network according to claim 1 is characterized in that: Step S5 includes the following steps: The joint learning objective of the causal adaptive recognition module is expressed as: ； in is the variational lower bound, as the optimization objective of joint learning, is a dynamic weight parameter that balances the weights of the likelihood term and the prior term, is the expected likelihood, which measures the ability of data to be reconstructed under the current causal graph. is the expected prior, reflecting the prior causal relationship in the knowledge graph, is the KL divergence of the causal edge e, which measures the difference between the variational distribution and the prior distribution, is the regularization weight of the causal edge e, which adjusts the regularization strength based on its importance score; The optimization process is as follows: Initialization: Initialize the causal graph G based on the knowledge graph and the prior probability matrix P(G); Variational Inference: In each iteration step, the variational lower bound is first optimized by variational inference ; Bayesian optimization: Secondly, Bayesian optimization is used to adjust the model hyperparameter θ; Dynamic update: According to the optimization results, the causal graph G and weight matrix W are dynamically updated to ensure the real-time and accuracy of the causal relationship.

Images