Loan risk assessment method and device, storage medium and electronic equipment

Abstract

The invention provides a loan risk assessment method and device, a storage medium and electronic equipment, and is applied to the technical field of artificial intelligence. According to the method, firstly, standardization processing is performed on loan application data, a pre-trained deep learning risk assessment model is input, and a loan risk score and hidden layer representation are obtained; and through the optimized orthogonal rotation matrix, mapping the hidden layer representation to a predefined loan risk key causal variable space, and outputting a corresponding causal variable. The variables are used for performing causal intervention verification, after an intervention result is obtained, the intervention result and a loan risk score are input into a visual interface together, a layered causal graph is generated, and a decision path of the risk score is visually displayed, so that the interpretability and transparency of loan risk assessment are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a loan risk assessment method, apparatus, storage medium, and electronic device. Background Technology

[0002] Currently, deep learning models have good predictive performance in loan risk assessment and can capture complex nonlinear relationships. However, their "black box" nature makes the decision-making process difficult to explain, affecting regulatory compliance and customer trust.

[0003] Existing technologies primarily rely on feature importance analysis or rule extraction methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations). While these methods improve transparency to some extent, they only reveal surface relationships and lack a deep understanding of the internal causal mechanisms of the model, making it difficult to verify model fairness and the dependence on explaining complex risk factors. Furthermore, while graph data aggregation and federated learning methods contribute to privacy protection and partial explanation, they do not address the systematic verification of causal paths.

[0004] Therefore, how to improve the interpretability of loan risk assessment has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides a loan risk assessment method, apparatus, storage medium, and electronic device that overcomes or at least partially solves the above problems. The technical solution is as follows:

[0006] A loan risk assessment method includes:

[0007] Obtain standardized feature data from loan application data;

[0008] The standardized feature data is input into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model.

[0009] The hidden layer representation is subjected to distributed alignment search: the hidden layer representation is mapped to the causal variable space composed of predefined key abstract concepts of loan risk using an optimized orthogonal rotation matrix, and the causal variable representation is output.

[0010] The causal variable representation was used to verify the causal intervention and obtain the intervention verification results.

[0011] The loan risk score and the intervention verification results are input into a visualization and interpretation interface to construct a hierarchical causal graph, which displays the assessment and decision-making path of the loan risk score.

[0012] Optionally, the standardized feature data for obtaining loan application data includes:

[0013] The loan application data is standardized and multidimensional features are extracted to generate standardized feature data.

[0014] Optionally, the standardization processing and multi-dimensional feature extraction of the loan application data to generate standardized feature data includes:

[0015] The structured data in the loan application data is normalized to obtain the first data to be processed;

[0016] Multimodal feature extraction is performed on the unstructured data in the loan application data to obtain the second data to be processed;

[0017] Using the first data to be processed and the second data to be processed, a comprehensive feature set including customer identity, contract, economic and behavioral dimensions is constructed.

[0018] Add time series features to the comprehensive feature set to generate standardized feature data.

[0019] Optionally, the step of mapping the hidden layer representation to a predefined causal variable space composed of key abstract concepts of loan risk using an optimized orthogonal rotation matrix, and outputting a causal variable representation, includes:

[0020] An optimization objective function for the orthogonal rotation matrix is ​​constructed to maximize the accuracy of the interchange intervention between the hidden layer representation and the defined key abstract concept of loan risk;

[0021] The orthogonal rotation matrix is ​​iteratively optimized using the gradient descent algorithm;

[0022] The hidden layer representation is mapped to the causal variable space using an optimized orthogonal rotation matrix, and the causal variable representation is output.

[0023] Optionally, the intervention verification results include counterfactual experiment verification data, risk factor dependence index data, sensitive variable verification data, equal opportunity verification index data, and average causal effect index data.

[0024] Optionally, the hierarchical causal graph is labeled with path weights and anonymized customer data.

[0025] Optionally, after inputting the loan risk score and the intervention verification results into the visualization and interpretation interface to construct a hierarchical causal graph, the method further includes:

[0026] In response to a user's adjustment of feature values ​​in the hierarchical causal graph, the hierarchical causal graph and the loan risk score are updated.

[0027] A loan risk assessment device includes: a standardized feature data acquisition unit, a hidden layer representation acquisition unit, a causal variable representation output unit, an intervention verification result acquisition unit, and a hierarchical causal graph construction unit.

[0028] The standardized feature data acquisition unit is used to acquire standardized feature data of loan application data;

[0029] The hidden layer representation acquisition unit is used to input the standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model.

[0030] The causal variable representation output unit is used to perform a distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix to map the hidden layer representation to a causal variable space composed of predefined key abstract concepts of loan risk, and outputting the causal variable representation;

[0031] The intervention verification result acquisition unit is used to perform causal intervention verification using the causal variable representation to obtain intervention verification results.

[0032] The hierarchical causal graph construction unit is used to input the loan risk score and the intervention verification result into the visualization and interpretation interface to construct a hierarchical causal graph, wherein the hierarchical causal graph displays the assessment and decision-making path of the loan risk score.

[0033] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the loan risk assessment method described above.

[0034] An electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the loan risk assessment method.

[0035] By employing the above technical solutions, this invention provides a loan risk assessment method, apparatus, storage medium, and electronic device. The method obtains standardized feature data from loan application data; inputs this standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model; performs a distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix, the hidden layer representation is mapped to a causal variable space composed of predefined key abstract concepts of loan risk, outputting causal variable representations; uses these causal variable representations to perform causal intervention verification, obtaining intervention verification results; and inputs the loan risk score and intervention verification results into a visualization interface to construct a hierarchical causal graph, where the hierarchical causal graph displays the assessment decision path of the loan risk score. This invention, by mapping the hidden layer representation of the deep learning model to a predefined key causal variable space of loan risk, and combining causal intervention verification with hierarchical causal graph visualization, effectively reveals the internal decision-making mechanism and causal relationships of the model, thereby significantly improving the transparency and interpretability of loan risk assessment results.

[0036] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0038] Figure 1 A flowchart illustrating one implementation of the loan risk assessment method provided in this invention is shown.

[0039] Figure 2 The diagram illustrates a specific implementation of the loan risk assessment method provided in this invention.

[0040] Figure 3 The diagram shows a specific implementation of step S200 in the loan risk assessment method provided by the present invention.

[0041] Figure 4 The diagram shows a specific implementation of step S120 in the loan risk assessment method provided by the present invention.

[0042] Figure 5A schematic diagram of the structure of the loan risk assessment device provided in an embodiment of the present invention is shown;

[0043] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0044] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0045] With the development of fintech, deep learning models have demonstrated outstanding predictive performance in loan risk assessment, effectively capturing complex nonlinear relationships within multi-dimensional customer information and improving the accuracy of risk identification. However, the inherent "black box" nature of deep learning models makes their internal decision-making logic difficult to understand and explain intuitively, posing a challenge for banks in explaining the reasons for loan approval to regulators and customers, thereby affecting the model's regulatory compliance and customer trust.

[0046] In existing technologies, mainstream interpretability methods mainly rely on feature importance analysis techniques, such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), or rule extraction methods such as decision tree paths. While these methods can reveal the impact of input features on model output to some extent, they are often limited to superficial statistical correlations and lack in-depth analysis of the model's internal causal mechanisms, making it difficult to explain how the model makes decisions based on key risk factors. Furthermore, some techniques improve model interpretability by aggregating features through graph data, but do not involve the analysis and verification of the model's causal mechanisms; risk control models under the federated learning framework, while enhancing data privacy protection, also lack effective verification of the model's internal causal paths.

[0047] Traditional methods fall short in explaining the interactions of complex risk factors and their contribution to model predictions, making it difficult to meet the stringent requirements of financial regulators regarding model transparency, fairness, and decision rationality. This makes it difficult for regulators to verify the fairness and compliance of models, and customers also struggle to effectively understand the basis of their decisions, hindering the widespread application of deep learning models in loan risk assessment.

[0048] Causal abstraction theory, through mechanism transformation and intervention algebra, provides a unified mathematical language and analytical tools for understanding the internal workings of deep learning models. It can map the internal representations of the model to an externally defined risk concept space, enabling a systematic understanding and verification of the model's causal structure. However, the specific application of causal abstraction theory in bank loan risk assessment has not yet been fully researched and practiced.

[0049] Based on this, this invention provides a loan risk assessment method. First, loan application data is standardized and input into a pre-trained deep learning risk assessment model to obtain a loan risk score and hidden layer representation. Through an optimized orthogonal rotation matrix, the hidden layer representation is mapped to a predefined space of key causal variables for loan risk, and the corresponding causal variables are output. These variables are used for causal intervention verification. After obtaining the intervention results, they are input into a visualization interface along with the loan risk score to generate a hierarchical causal graph, intuitively displaying the decision path of the risk score, thereby improving the interpretability and transparency of loan risk assessment. It is evident that this invention overcomes the limitations of existing feature importance analysis, aligns the internal representation of the deep neural network with causal concepts related to loan risk, constructs a causal intervention verification mechanism, quantifies the model's dependence on key risk factors, and enhances the transparency and interactivity of model explanation through dynamic visualization and counterfactual simulation, thereby meeting regulatory compliance requirements, increasing customer trust, and promoting continuous model optimization.

[0050] like Figure 1 The diagram shows a flowchart of one embodiment of the loan risk assessment method provided by this invention. The method may include:

[0051] S100, Standardized feature data for obtaining loan application data.

[0052] Loan application data refers to all relevant information submitted by individuals or businesses when applying for a loan, including various types of data such as income, assets, credit history, and occupation.

[0053] Standardized feature data refers to the feature data obtained by using standardization methods (such as Z-score standardization or Min-Max standardization) after preprocessing loan application data.

[0054] Specifically, embodiments of the present invention can extract relevant loan application data from a loan application database, including the borrower's personal information, financial status, and credit history, and then use standardization techniques to process these features to make them uniform on the same scale. The standardization process may include calculating the mean and standard deviation of the features, and converting the original data into standardized data using formulas to improve the efficiency and accuracy of subsequent model training.

[0055] S110. Input the standardized feature data into the pre-trained deep learning risk assessment model to obtain the loan risk score and the corresponding hidden layer representation output by the deep learning risk assessment model.

[0056] Deep learning risk assessment models refer to models trained using deep learning algorithms such as deep neural networks, used to assess the default risk of loan applicants. These models can learn complex risk assessment patterns through training on historical loan data.

[0057] Loan risk score refers to the quantitative assessment of a loan applicant's credit risk by a deep learning risk assessment model, representing the probability or risk level of the applicant's default. The loan risk score can be presented as a numerical score; a higher score indicates greater risk, and vice versa.

[0058] Hidden layer representation refers to the abstract expression of input features in the hidden layer of a deep learning model. It is an intermediate result after the deep learning model performs complex transformations on the data features.

[0059] Specifically, in this embodiment of the invention, standardized feature data can be input into a pre-trained deep learning risk assessment model, which can be a fully connected neural network or a convolutional neural network. The deep learning risk assessment model can calculate a loan risk score through forward propagation and generate the corresponding hidden layer representation in the hidden layer output. The output of the deep learning risk assessment model includes not only the loan risk score but also the intermediate feature representations learned by the model during the calculation process, i.e., the hidden layer representations.

[0060] S120. Perform distributed alignment search on the hidden layer representation: use an optimized orthogonal rotation matrix to map the hidden layer representation to the causal variable space composed of predefined key abstract concepts of loan risk, and output the causal variable representation.

[0061] Distributed Alignment Search (DAS) is an interpretability method based on causal abstraction theory, used to align hidden layer representations to a predefined space of causal variables. In this way, complex representations can be mapped to more interpretable abstract concepts.

[0062] Among them, the orthogonal rotation matrix refers to a linear transformation matrix that preserves the length and angle of a vector. It is used to rotate data in space, preserve the distance and similarity features of the data, transform the data from one coordinate space to another, and realize the mapping between the hidden layer representation and the causal variable space.

[0063] Among them, the key abstract concept of loan risk refers to the core factors that affect loan risk, such as repayment ability, fraud risk and economic stability, which are key abstract variables for risk assessment.

[0064] The causal variable space refers to a multidimensional space composed of key abstract concepts of loan risk, used to represent the causal relationships and effects between various factors, and to analyze and understand the changes in loan risk under different circumstances.

[0065] Among them, the causal variable representation refers to the expression result obtained after transforming the hidden layer representation to the causal variable space, which reflects the specific values ​​and relationships of each causal factor.

[0066] Specifically, embodiments of the present invention can maximize the causal alignment between the hidden layer representation and the preset loan risk abstraction concepts (including repayment ability, fraud risk, and economic stability) by constructing an optimization objective function for the orthogonal rotation matrix; the optimal rotation parameters are iteratively solved using the gradient descent algorithm to map the original hidden layer representation to the causal variable space composed of the above-mentioned risk concepts, and finally output variable representations that conform to the causal assumptions.

[0067] S130. Use causal variables to verify causal intervention and obtain the intervention verification results.

[0068] Causal intervention validation refers to the process of assessing the changes in loan risk after intervention in causal variables under specific circumstances, and is used to test the model's response to and explanatory power regarding causal relationships.

[0069] The intervention validation results refer to the results obtained after causal intervention validation, which show the changes in loan risk scores after a specific intervention and are used to evaluate the accuracy and reliability of the model's causal inference.

[0070] Specifically, embodiments of the present invention can design and implement a series of intervention measures based on the generated causal variable representations to examine the impact of these measures on loan risk. For example, different income or borrowing conditions can be simulated, and the impact of interventions on causal variables on loan risk scores can be analyzed to obtain intervention validation results, thereby assessing which variables' changes will significantly reduce or increase risk.

[0071] S140. Input the loan risk score and intervention verification results into the visualization interpretation interface to construct a hierarchical causal graph, in which the hierarchical causal graph shows the assessment decision path with the loan risk score.

[0072] Among them, the visual explanation interface is a visualization platform used to display loan risk scores, causal variable representations, and intervention results. It can graphically display the model's decision-making process and its results, helping users understand the model's decision-making process.

[0073] Hierarchical cause-effect graphs are a graphical representation method that uses a hierarchical structure to show the relationships between different causal variables and the assessment decision-making path for loan risk scoring. Hierarchical cause-effect graphs help users more intuitively understand the logic and influencing factors of risk assessment.

[0074] Specifically, embodiments of the present invention can integrate loan risk scoring and intervention verification results into a single visual explanation interface. This interface constructs a hierarchical causal graph to display the relationships between different causal variables and the decision-making path for assessing loan risk scoring, enabling users to intuitively understand the model's decision-making process and influencing factors.

[0075] This invention provides a loan risk assessment method, which includes: obtaining standardized feature data of loan application data; inputting the standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model; performing a distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix to map the hidden layer representation to a causal variable space composed of predefined key abstract concepts of loan risk, and outputting causal variable representation; using the causal variable representation to perform causal intervention verification, and obtaining intervention verification results; inputting the loan risk score and intervention verification results into a visualization interpretation interface to construct a hierarchical causal graph, wherein the hierarchical causal graph displays the assessment decision path of the loan risk score. This invention effectively reveals the internal decision-making mechanism and causal relationships of the model by mapping the hidden layer representation of the deep learning model to a predefined key causal variable space of loan risk, and combining causal intervention verification with hierarchical causal graph visualization, thereby significantly improving the transparency and interpretability of loan risk assessment results.

[0076] Optional, based on Figure 1 The method shown is as follows: Figure 2 As shown, this is a flowchart illustrating a specific implementation of the loan risk assessment method provided in this embodiment of the invention. Step S100 may include:

[0077] S200: Standardize loan application data and extract multi-dimensional features to generate standardized feature data.

[0078] Specifically, embodiments of the present invention can standardize loan application data, including normalizing structured data to standardize numerical features, and extracting multimodal features from unstructured data to convert text and speech information. The processed data is then integrated to construct a feature set encompassing four dimensions: customer identity, contractual obligations, economic status, and behavior. Time-series features are embedded in each dimension to capture dynamic behavioral changes, thereby generating standardized feature data for final risk assessment.

[0079] This invention, through standardized processing and multi-dimensional feature extraction of loan application data, effectively eliminates dimensional differences and distribution biases among different features, making the data more easily understood and utilized by deep learning risk assessment models under a unified scale. Simultaneously, multi-dimensional feature extraction integrates structured and unstructured information, comprehensively reflecting the customer's identity, economic status, and behavioral characteristics, significantly improving the model's ability to discriminate loan risks and its robustness. This ensures the accuracy of subsequent loan risk scoring and the effectiveness of hidden layer representations, providing a reliable basis for causal variable mapping and intervention verification based on hidden layer representations, thereby comprehensively improving the scientific rigor and interpretability of loan risk assessment.

[0080] Optional, based on Figure 2 The method shown is as follows: Figure 3 The diagram shows a specific implementation of step S200 in the loan risk assessment method provided by this invention. Step S200 may include:

[0081] S300: Normalize the structured data in the loan application data to obtain the first data to be processed.

[0082] Structured data refers to numerical or categorical data in loan application data that is in a fixed format and is easy to store and analyze, such as the applicant's income, credit score, debt ratio, etc.

[0083] Specifically, in this embodiment of the invention, the structured data in loan application data is normalized to eliminate the dimensional differences between different indicators, so that each feature data is within a uniform numerical range, which facilitates the stable training and accurate analysis of the subsequent model, thereby obtaining the first data to be processed that is suitable for the input of the deep learning model.

[0084] S310. Perform multimodal feature extraction on the unstructured data in the loan application data to obtain the second data to be processed.

[0085] Unstructured data refers to data types in loan application data that do not have a fixed format or explicit fields, such as free text descriptions or voice recordings provided by customers.

[0086] Multimodal feature extraction refers to the process of extracting effective features from different modalities such as text and speech using technologies such as natural language processing and speech signal processing for unstructured data, and then transforming them into structured numerical vectors to facilitate model understanding and fusion.

[0087] Specifically, in this embodiment of the invention, for unstructured data in loan application data, a multimodal feature extraction technique is used. Through natural language processing, speech signal processing and other methods, unstructured information is transformed into structured numerical feature vectors, thereby capturing the customer's subjective intentions and emotional expressions, supplementing the deficiencies of structured data, and finally generating a second set of data to be processed with rich semantic information.

[0088] S320. Using the first and second pending data, construct a comprehensive feature set that includes customer identity, treaty, economic and behavioral dimensions.

[0089] Specifically, embodiments of the present invention can fuse structured normalized data and unstructured multimodal feature data, classify and organize them according to four dimensions: customer identity (such as age, occupation), terms (loan contract terms), economic status (income, debt, etc.), and behavioral patterns (consumption habits, repayment records), and construct a multi-dimensional feature set, thereby comprehensively and meticulously depicting the overall profile of the loan applicant and improving the granularity and accuracy of risk assessment.

[0090] S330. Add time series features to the comprehensive feature set to generate standardized feature data.

[0091] Among them, time series features refer to dynamic change features extracted based on customer historical data (such as repayment records and consumption behavior over the past few months), reflecting the trends and patterns of customer identity, contracts, economics, and behavior over time, and enhancing the model's ability to perceive changes in risk.

[0092] This invention, targeting a comprehensive feature set, further incorporates time series analysis methods to include dynamic change information such as repayment records and consumption patterns from the past six months. Time series features can reflect trends and fluctuations in customer behavior and economic conditions, providing temporal context for risk assessment models and enhancing their sensitivity to and predictive ability regarding risk changes. After time series enhancement and standardized processing, standardized feature data is generated for training and inference of deep learning models.

[0093] This invention, through standardized processing and multi-dimensional feature extraction of loan application data, achieves normalization of structured data and multi-modal feature extraction of unstructured data, ensuring a unified scale and rich representation of the data. Subsequently, it integrates a comprehensive feature set encompassing four dimensions: customer identity, contract, economic status, and behavior, comprehensively capturing multifaceted information about loan applicants. Furthermore, it incorporates time-series features to introduce dynamic changes in customer behavior, thereby generating high-quality standardized feature data. This not only improves the expressive power and model adaptability of the input data but also provides an accurate, comprehensive, and time-series-based input foundation for deep learning risk assessment models. It significantly enhances the accuracy of loan risk scoring and the interpretability of hidden layer representations, ensuring the effectiveness of causal variable mapping and intervention verification.

[0094] Optional, based on Figure 1 The method shown is as follows: Figure 4 The diagram shows a specific implementation of step S120 in the loan risk assessment method provided by this invention. Step S120 may include:

[0095] S400. Construct an optimization objective function for the orthogonal rotation matrix to maximize the accuracy of the interchange intervention between the hidden layer representation and the defined key abstract concept of loan risk.

[0096] Specifically, embodiments of the present invention can design an objective function and adjust the parameters of the rotation matrix so that the rotated hidden layer representation can accurately correspond to predefined key abstract concepts of loan risk (such as "repayment ability," "fraud risk," and "economic stability") in a specific dimension. The objective function is optimized to maximize the Interchange Intervention Accuracy (IIA), that is, when the value of a certain dimension variable is fixed, the model output should conform to the expected causal relationship, thereby ensuring that the rotation matrix can achieve effective alignment between the hidden representation and the causal concept.

[0097] The embodiments of the present invention provide for maximizing the accuracy of interchange interventions. The calculation formula is:

[0098]

[0099] The actual accuracy rate refers to the accuracy rate achieved by the model based on the rotated hidden layer representation during the mutual intervention verification process, reflecting the model's true performance in capturing the correspondence between causal variables and key abstract concepts of loan risk.

[0100] The virtual classifier accuracy represents a baseline accuracy, which can be the accuracy of random guessing or the simplest prediction strategy, and is used to measure the performance level when there is no actual causal alignment effect.

[0101] Task performance refers to the theoretical highest accuracy or optimal performance level of the intervention verification task, that is, the accuracy that the model can achieve under ideal conditions.

[0102] S410. Iteratively optimize the orthogonal rotation matrix using the gradient descent algorithm.

[0103] Specifically, embodiments of the present invention can iteratively adjust the parameters of the orthogonal rotation matrix using a gradient descent algorithm based on a constructed optimization objective function. In each iteration, the gradient of the objective function is calculated, and the rotation matrix parameters are updated along the gradient descent direction while maintaining the orthogonality of the matrix. By continuously optimizing the mapping effect, the correspondence between the hidden layer representation and the key concept space of loan risk is gradually improved, achieving the structuring and interpretability of the model's internal representation.

[0104] S420. Map the hidden layer representation to the causal variable space using the optimized orthogonal rotation matrix, and output the causal variable representation.

[0105] Specifically, in this embodiment of the invention, the hidden layer representation of the deep learning model can be multiplied by the orthogonal rotation matrix to achieve a mapping transformation from the original feature space to the predefined causal variable space, thereby obtaining a causal variable representation. This transforms the complex hidden layer activation values ​​into variables with clear semantic and causal meanings, which is helpful for subsequent causal intervention verification and risk interpretation.

[0106] This invention maximizes the accuracy of intervention by interchange between the hidden layer representation and predefined key abstract concepts of loan risk by constructing an optimization objective function for an orthogonal rotation matrix. The rotation matrix is ​​iteratively optimized using a gradient descent algorithm to ensure precise alignment of the hidden layer representation in the causal variable space. Applying the optimized orthogonal rotation matrix to the mapping of the hidden layer representation not only improves the semantic clarity and interpretability of the causal variable representation but also enhances the model's ability to capture causal relationships, effectively supporting subsequent causal intervention verification and significantly improving the transparency and reliability of risk scoring.

[0107] Optionally, the intervention verification results provided in the embodiments of the present invention include counterfactual experiment verification data, risk factor dependence index data, sensitive variable verification data, equal chance verification index data, and average causal effect index data.

[0108] Counterfactual experiment validation data refers to experimental data designed based on counterfactual assumptions. This data is obtained by artificially modifying the values ​​of a specific dimension (corresponding to a key factor in loan risk) in the rotated hidden layer representation (e.g., assuming a 20% increase in customer income), and observing the changes in the loan risk score output by the model. Counterfactual experiment validation data is used to verify whether the model responds to expected variable changes at the causal level, reflecting the effectiveness of the model's causal inference.

[0109] The Risk Factor Dependence (RFD) metric quantifies the model's dependence on various loan risk factors. By analyzing the model's output sensitivity to changes in the dimensions of causal variables after rotation, the RFD reflects the importance and influence of different risk factors in the model's decision-making, helping to identify the core risk factors emphasized by the model. A higher RFD value indicates a greater dependence of the model on that risk factor, and that this dependence aligns with the model's causal structure assumptions. Conversely, a lower RFD value indicates a weaker influence of the risk factor or a mismatch with the causal assumptions.

[0110] The risk factor dependence calculation formula provided in this embodiment of the invention is as follows:

[0111]

[0112] In this context, the post-intervention risk score change refers to the change in the model's output risk score when a dimension corresponding to a specific risk factor in the rotated hidden layer is modified (e.g., increasing customer revenue by 20%) during a causal intervention experiment. The post-intervention risk score change reflects the direct impact of this risk factor change on the overall risk score. The original risk score is the risk score calculated by the model for the same input sample without any intervention. The original risk score serves as a baseline value to normalize the risk score change and avoid the influence of differences in scoring scales. The larger the ratio of the post-intervention risk score change to the original risk score, the stronger the model's dependence on the corresponding risk factor. The IIA matching degree is used to adjust the reliability of the impact of the risk factor change on the model score, ensuring that the measurement results conform to a reasonable causal structure and the principle of choice independence.

[0113] The sensitive variable validation data is used to verify the weight and effect of the model on sensitive variables (such as gender, race, etc.) and to assess whether there is any potential bias or discrimination. By measuring the difference in risk scores caused by changes in sensitive variables in the model output while controlling for other variables, the sensitive variable validation data provides direct evidence for the fairness of the model.

[0114] The equality of opportunity validation index measures the degree to which different sensitive groups have the same chance of obtaining loan approval under the same risk level. This index reflects whether the model maintains consistent positive and negative decision probabilities across different groups, thus indicating whether the model satisfies the principle of equality of opportunity in financial fairness.

[0115] Among them, the average causal effect index data is used to quantify the average causal impact of a specific variable on loan risk scoring. By analyzing the average difference in the output of the statistical model when the variable takes different values ​​after controlling for other factors, the average causal effect index data reveals the true causal role of the variable in the model's decision-making, and helps to assess and control potential biases in the model.

[0116] This invention validates causal interventions based on causal variable representations. The resulting validation data encompasses counterfactual experiment data, risk factor dependence index data, sensitive variable validation data, equal chance validation index data, and average causal effect index data. These multi-dimensional validation results comprehensively reveal the model's decision-making mechanism, quantify the influence of different risk factors, assess the model's fairness and causal rationality, and effectively improve the interpretability and credibility of the risk assessment model.

[0117] Optionally, the hierarchical causal graph provided in this embodiment of the invention is labeled with path weights and anonymized customer data.

[0118] The causal graph displayed in the visual explanation interface provided in this embodiment of the invention not only clearly presents the hierarchical causal relationships between various causal variables, but also intuitively expresses the contribution of different causal paths to the final loan risk score through path weights, helping users understand which factors play a key role in the model's decision-making. At the same time, the customer's specific sensitive information is anonymized and displayed only in the form of abstract causal variables and path weights, thus protecting customer privacy while ensuring the transparency and interpretability of causal reasoning.

[0119] Optionally, in the above Figure 1 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, after step S140, the method may further include:

[0120] In response to user adjustments to the feature values ​​in the hierarchical causal graph, update the hierarchical causal graph and the loan risk score.

[0121] The visual explanation interface provided in this invention supports dynamic interactive functions, allowing users to modify input feature values ​​in real time (e.g., adjusting customer income level, credit history, etc.) through dragging or slider operations. Based on the latest input, the system automatically recalculates variable values ​​and path weights on the causal path and updates the risk score in real time, thereby displaying the adjusted counterfactual results. This dynamic feedback mechanism enables users to intuitively perceive the impact of changes in various features on loan risk, which is helpful for risk management, decision optimization, and strategy simulation. Users can also switch from macro-level risk concepts (such as "fraud risk") to micro-level features (such as "recent frequent job changes"), enhancing their understanding of the bank's decision-making process.

[0122] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0123] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0124] Corresponding to the above method embodiments, this invention also provides a loan risk assessment device, the structure of which is as follows: Figure 5 As shown, it may include: a standardized feature data acquisition unit 10, a hidden layer representation acquisition unit 20, a causal variable representation output unit 30, an intervention verification result acquisition unit 40, and a hierarchical causal graph construction unit 50.

[0125] Standardized feature data acquisition unit 10 is used to acquire standardized feature data of loan application data.

[0126] The hidden layer representation acquisition unit 20 is used to input standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model.

[0127] The causal variable representation output unit 30 is used to perform distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix, the hidden layer representation is mapped to the causal variable space composed of predefined key abstract concepts of loan risk, and the causal variable representation is output.

[0128] The intervention validation result acquisition unit 40 is used to perform causal intervention validation using causal variable representation and obtain intervention validation results.

[0129] The hierarchical causal graph construction unit 50 is used to input loan risk scores and intervention verification results into the visualization and interpretation interface to construct a hierarchical causal graph, which displays the assessment and decision-making path of loan risk scores.

[0130] Optionally, the standardized feature data acquisition unit 10 is specifically used to standardize loan application data and extract multi-dimensional features to generate standardized feature data.

[0131] Optionally, the standardized feature data acquisition unit 10 may include: a first data to be processed acquisition subunit, a second data to be processed acquisition subunit, a comprehensive feature set construction subunit, and a standardized feature data generation subunit.

[0132] Optionally, the causal variable representation output unit 30 can be specifically used to construct the optimization objective function of the orthogonal rotation matrix to maximize the accuracy of the interchange intervention between the hidden layer representation and the defined key abstract concept of loan risk; iteratively optimize the orthogonal rotation matrix using the gradient descent algorithm; map the hidden layer representation to the causal variable space through the optimized orthogonal rotation matrix, and output the causal variable representation.

[0133] Optionally, the intervention validation results include counterfactual experiment validation data, risk factor dependence index data, sensitive variable validation data, equal opportunity validation index data, and average causal effect index data.

[0134] Optionally, the hierarchical causal graph is labeled with path weights and anonymized customer data.

[0135] Optionally, the loan risk assessment device may also include an operation response unit.

[0136] The operation response unit is used by the hierarchical causal graph construction unit 50 to input the loan risk score and intervention verification results into the visualization interpretation interface. After constructing the hierarchical causal graph, it updates the hierarchical causal graph and the loan risk score in response to the user's adjustment operation on the feature values ​​in the hierarchical causal graph.

[0137] This invention provides a loan risk assessment device, which is used to: obtain standardized feature data of loan application data; input the standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model; perform distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix to map the hidden layer representation to a causal variable space composed of predefined key abstract concepts of loan risk, and output causal variable representation; use the causal variable representation to perform causal intervention verification to obtain intervention verification results; input the loan risk score and intervention verification results into a visualization interpretation interface to construct a hierarchical causal graph, wherein the hierarchical causal graph displays the assessment decision path of the loan risk score. This invention effectively reveals the internal decision-making mechanism and causal relationships of the model by mapping the hidden layer representation of the deep learning model to a predefined key causal variable space of loan risk, and combines causal intervention verification with hierarchical causal graph visualization, thereby significantly improving the transparency and interpretability of loan risk assessment results.

[0138] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0139] The loan risk assessment device includes a processor and a memory. The standardized feature data acquisition unit 10, the hidden layer representation acquisition unit 20, the causal variable representation output unit 30, the intervention verification result acquisition unit 40, and the hierarchical causal graph construction unit 50 are all stored in the memory as program units. The processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0140] The processor contains kernels that retrieve corresponding program units from memory. One or more kernels can be configured, and kernel parameters are adjusted to standardize loan application data. This data is then input into a pre-trained deep learning risk assessment model to obtain a loan risk score and hidden layer representation. An optimized orthogonal rotation matrix maps the hidden layer representation to a predefined space of key causal variables for loan risk, outputting the corresponding causal variables. These variables are used for causal intervention verification. The intervention results, along with the loan risk score, are input into a visualization interface to generate a hierarchical causal graph. This graph visually displays the decision path of the risk score, thereby improving the interpretability and transparency of loan risk assessment.

[0141] This invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the loan risk assessment method.

[0142] This invention provides a processor for running a program, wherein the program executes the loan risk assessment method during runtime.

[0143] like Figure 6 As shown, this embodiment of the invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003. The processor 1001 and the memory 1002 communicate with each other via the bus 1003. The processor 1001 is used to call program instructions in the memory 1002 to execute the aforementioned loan risk assessment method. The electronic device in this document can be a server, PC, PAD, mobile phone, etc.

[0144] The present invention also provides a computer program product that, when executed on an electronic device, is suitable for executing a program that initializes a loan risk assessment method step.

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] In a typical configuration, an electronic device includes one or more processors (CPUs), memory, and a bus. The electronic device may also include input / output interfaces, network interfaces, etc.

[0147] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0148] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0149] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0150] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0151] In the description of this invention, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0152] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0153] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the present invention.

Claims

1. A loan risk assessment method, characterized in that, include: Obtain standardized feature data from loan application data; The standardized feature data is input into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model. The hidden layer representation is subjected to distributed alignment search: the hidden layer representation is mapped to the causal variable space composed of predefined key abstract concepts of loan risk using an optimized orthogonal rotation matrix, and the causal variable representation is output. The causal variable representation was used to verify the causal intervention and obtain the intervention verification results. The loan risk score and the intervention verification results are input into a visualization and interpretation interface to construct a hierarchical causal graph, which displays the assessment and decision-making path of the loan risk score.

2. The method according to claim 1, characterized in that, The standardized feature data for obtaining loan application data includes: The loan application data is standardized and multidimensional features are extracted to generate standardized feature data.

3. The method according to claim 2, characterized in that, The standardization and multi-dimensional feature extraction of the loan application data to generate standardized feature data includes: The structured data in the loan application data is normalized to obtain the first data to be processed; Multimodal feature extraction is performed on the unstructured data in the loan application data to obtain the second data to be processed; Using the first data to be processed and the second data to be processed, a comprehensive feature set including customer identity, contract, economic and behavioral dimensions is constructed. Add time series features to the comprehensive feature set to generate standardized feature data.

4. The method according to claim 1, characterized in that, The process involves using an optimized orthogonal rotation matrix to map the hidden layer representation to a causal variable space comprised of predefined key abstract concepts of loan risk, outputting a causal variable representation, including: An optimization objective function for the orthogonal rotation matrix is ​​constructed to maximize the accuracy of the interchange intervention between the hidden layer representation and the defined key abstract concept of loan risk; The orthogonal rotation matrix is ​​iteratively optimized using the gradient descent algorithm; The hidden layer representation is mapped to the causal variable space using an optimized orthogonal rotation matrix, and the causal variable representation is output.

5. The method according to claim 1, characterized in that, The intervention validation results include counterfactual experiment validation data, risk factor dependence index data, sensitive variable validation data, equal opportunity validation index data, and average causal effect index data.

6. The method according to claim 1, characterized in that, The hierarchical causal graph is labeled with path weights and anonymized customer data.

7. The method according to claim 1, characterized in that, After inputting the loan risk score and the intervention verification results into the visualization and interpretation interface to construct a hierarchical causal graph, the method further includes: In response to a user's adjustment of feature values ​​in the hierarchical causal graph, the hierarchical causal graph and the loan risk score are updated.

8. A loan risk assessment device, characterized in that, include: The system includes a standardized feature data acquisition unit, a hidden layer representation acquisition unit, a causal variable representation output unit, an intervention validation result acquisition unit, and a hierarchical causal graph construction unit. The standardized feature data acquisition unit is used to acquire standardized feature data of loan application data; The hidden layer representation acquisition unit is used to input the standardized feature data into a pre-trained deep learning risk assessment model to obtain the loan risk score and corresponding hidden layer representation output by the deep learning risk assessment model. The causal variable representation output unit is used to perform a distributed alignment search on the hidden layer representation: using an optimized orthogonal rotation matrix to map the hidden layer representation to a causal variable space composed of predefined key abstract concepts of loan risk, and outputting the causal variable representation; The intervention verification result acquisition unit is used to perform causal intervention verification using the causal variable representation to obtain intervention verification results. The hierarchical causal graph construction unit is used to input the loan risk score and the intervention verification result into the visualization and interpretation interface to construct a hierarchical causal graph, wherein the hierarchical causal graph displays the assessment and decision-making path of the loan risk score.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the loan risk assessment method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the loan risk assessment method as described in any one of claims 1 to 7.

Images