Credit risk attribution method and apparatus

Abstract

Embodiments of the present specification disclose a credit risk attribution method and device. The method is executed by a computer system and includes the following steps: determining a credit user group to be analyzed according to abnormal fluctuations of a credit risk indicator in a credit business; performing dimension reduction processing on an original feature set related to the group to obtain a key feature subset; then, constructing a labeled data set based on feature values of each user in the group on the key feature subset and a corresponding risk label, wherein the risk label indicates whether the user is in a preset risk state defined according to the credit risk indicator; thereafter, determining a decision tree path meeting a predetermined standard based on the labeled data set, wherein the path includes feature items and corresponding value conditions, and the predetermined standard includes that a coverage rate of the decision tree path for users in the preset risk state reaches a predetermined threshold; and then, generating an attribution conclusion of the abnormal fluctuations based on the decision tree path.

Description

Technical Field

[0001] This specification relates to the field of machine learning technology, and more particularly to a credit risk attribution method and apparatus, a computer-readable storage medium, and a computing device. Background Technology

[0002] In modern computing systems, especially in credit service systems that require handling large-scale user credit assessments and automated decision-making, maintaining the stability and interpretability of platform decisions is crucial. However, risk indicators in credit operations, such as customer delinquency rates, often experience unexpected and abnormal fluctuations. Accurately and quickly identifying the root causes of these risk indicator fluctuations is a core technical challenge for optimizing the system's decision-making logic, ensuring service stability, and maintaining a healthy user ecosystem within the system.

[0003] Currently, attribution analysis for such risk fluctuations mainly relies on manual operation by analysts. This traditional model has significant limitations at the technical level, restricting the platform's analytical capabilities and response efficiency, specifically: 1) Low efficiency of manual attribution: Strategy analysts need to manually pull data and analyze it dimension by dimension, and a complete attribution often takes several days; 2) Large amount of repetitive workload: A similar attribution process needs to be repeated for each risk fluctuation.

[0004] In summary, existing technical solutions have significant shortcomings in attribution analysis of fluctuations in credit risk indicators, making it difficult to achieve high efficiency, accuracy, and automation. Summary of the Invention

[0005] This specification describes a credit risk attribution method and apparatus that can solve the above-mentioned technical problems.

[0006] According to the first aspect, a credit risk attribution method is provided, executed by a computer system. The method includes the following steps:

[0007] Based on abnormal fluctuations in credit risk indicators in credit operations, a group of credit users to be analyzed is identified. The original feature set involving this user group is subjected to dimensionality reduction processing to obtain a key feature subset. Based on the feature values ​​of each user in the credit user group on this key feature subset, and their corresponding risk labels, a labeled dataset is constructed; the risk labels indicate whether a user is in a preset risk state, the definition of which depends on the credit risk indicators. Based on the labeled dataset, a decision tree path conforming to predetermined criteria is determined; the decision tree path includes feature terms and corresponding value conditions, the predetermined criteria including: the coverage rate of the decision tree path for users in the preset risk state reaches a predetermined threshold. Based on the decision tree path, attribution conclusions for the abnormal fluctuations are generated.

[0008] In one embodiment, the dimensionality reduction process includes: matching the original feature set with a risk feature set in an expert knowledge base; the risk feature set corresponds to the credit business customer groups affected by historical macroeconomic events; and, if the matching is successful, constructing the key feature subset using the matched features.

[0009] Furthermore, in a specific embodiment, the dimensionality reduction process further includes: in the case of matching failure, performing principal component analysis on the feature data of each user in the credit user group under the original feature set to obtain selected principal components; selecting the top k original features that contribute the most to the principal components to form the key feature subset.

[0010] In one embodiment, determining a decision tree path that meets predetermined criteria based on the labeled dataset includes: calculating the discrimination ability index of each key feature against the risk label based on the labeled dataset, thereby selecting the key feature with the strongest discrimination ability and its corresponding feature value conditions as the first-level node of the decision tree path; calculating the discrimination ability index of the remaining key features against the risk label in the labeled data subset defined by the first-level node, thereby selecting the key feature with the strongest discrimination ability and its corresponding feature value conditions as the second-level node of the decision tree path; recursively executing the above steps until the generated decision tree path meets the predetermined criteria.

[0011] Furthermore, in a specific embodiment, the method further includes: if the depth of the generated decision tree path has reached a preset depth limit, but still has not reached the predetermined standard, then one or more additional decision tree paths are generated until the predetermined standard is reached.

[0012] In one embodiment, the abnormal fluctuations involve an observation period and a baseline period for the credit risk indicator. The method further includes: dividing each characteristic value of a predetermined characteristic item of the credit user group into bins, and calculating the structural proportion and risk performance value of each bin in each period of the baseline period and the observation period; wherein, the structural proportion refers to the proportion of the business scale metric of the bin in the total business scale within the corresponding period, and the risk performance value refers to the proportion of the number of users in the bin who triggered the credit risk indicator to the total number of users in the bin. Based on the calculated structural proportion and risk performance value, the structural effect contribution value caused by changes in the user group structure and the point estimation effect contribution value caused by changes in the user group quality are further calculated. The structural effect contribution value and the point estimation effect contribution value are used to generate the attribution conclusion for the abnormal fluctuations.

[0013] Furthermore, in a specific embodiment, there are multiple predetermined feature items; the method further includes: determining the total contribution value of structural change based on the structural effect contribution value of each predetermined feature item; and determining the total contribution value of quality change based on the point estimation effect contribution value of each predetermined feature item. The total contribution value of structural change is compared with the total contribution value of quality change, and a dominant feature item is determined from the multiple predetermined feature items based on the comparison result. Target bins are selected from the dominant feature items based on the risk performance values ​​of each feature value bin under the dominant feature item. The user group defined by the target bins is taken as the object to be analyzed. From the remaining feature items after excluding the dominant feature item, the structural proportion and risk performance values ​​of each feature value bin under each predetermined feature item are re-obtained in the baseline period and observation period, and the structural effect contribution value and point estimation effect contribution value are calculated accordingly for recursive analysis.

[0014] In one embodiment, the abnormal fluctuations involve an observation period and a benchmark period for the credit risk indicators; the method further includes: obtaining the overall risk indicator values ​​and the overall risk indicator calculation base for each period within the benchmark period and the observation period, and obtaining the risk indicator values ​​and risk indicator calculation bases for the target user group defined by target feature value binning of predetermined feature items for each period. Based on the obtained data, the total contribution ratio and the size ratio of the target user group are calculated; wherein, the total contribution ratio quantifies the degree of contribution of the group's own risk level changes and size structure changes to the abnormal fluctuations, and the size ratio is the proportion of the group's risk indicator calculation base to the overall risk indicator calculation base in the observation period. The ratio between the total contribution ratio and the size ratio is calculated as the attention coefficient of the target user group; wherein, the attention coefficient is used to generate the attribution conclusion of the abnormal fluctuations.

[0015] In one embodiment, the abnormal fluctuations involve an observation period and a baseline period for the credit risk indicator, during which a target business strategy was implemented. The method further includes: determining a test group and a control group corresponding to the target business strategy, wherein the test group is a subgroup of users to whom the target business strategy was applied during the observation period, and the control group is a subgroup of users to whom the target business strategy was not applied during the same period. Based on data from the test group and the control group during the baseline period before the implementation of the target business strategy, a homogeneity test is performed on the test group and the control group. Provided the homogeneity test passes, based on data from the test group and the control group during the observation period after the implementation of the target business strategy, the true contribution value of the target business strategy to the business indicator is calculated; wherein the true contribution value is used to generate the attribution conclusion for the abnormal fluctuations.

[0016] Furthermore, in one specific embodiment, the method further includes: in the absence of a control group corresponding to the target business strategy, matching similar historical strategies of the target business strategy from a historical strategy library, and determining the control group corresponding to the similar historical strategy as a substitute control group. Based on the data of the test group during the observation period after the implementation of the target business strategy, and the data of the substitute control group during the historical observation period after the implementation of the similar historical strategy, calculating the estimated contribution value of the target business strategy to the business indicators. The estimated contribution value is used to generate the attribution conclusion for the abnormal fluctuations.

[0017] According to the second aspect, a credit risk attribution device is provided, integrated into a computer system. The device includes the following functional units:

[0018] The analysis group identification unit is configured to identify the credit user group to be analyzed based on abnormal fluctuations in credit risk indicators in credit business. The dimensionality reduction processing unit is configured to perform dimensionality reduction processing on the original feature set involved in the credit user group to obtain a key feature subset. The labeled data construction unit is configured to construct a labeled dataset based on the feature values ​​of each user in the credit user group on the key feature subset and the corresponding risk labels; the risk labels indicate whether a user is in a preset risk state, and the definition of the preset risk state depends on the credit risk indicators. The path determination unit is configured to determine a decision tree path that meets predetermined criteria based on the labeled dataset; the decision tree path includes feature items and corresponding value conditions, and the predetermined criteria include: the coverage rate of the decision tree path for users in the preset risk state reaches a predetermined threshold. The conclusion generation unit is configured to generate attribution conclusions for the abnormal fluctuations based on the decision tree path.

[0019] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in the first aspect.

[0020] According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method provided in the first aspect.

[0021] In summary, the methods and apparatus disclosed in the embodiments of this specification can improve the automation and intelligence of credit risk attribution, achieve refined root cause localization, enhance the stability of system or platform risk control operations, optimize resource allocation, and ultimately protect the user experience and rights in the credit service system. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This diagram illustrates the implementation architecture of the credit risk attribution scheme disclosed in the embodiments of this specification;

[0024] Figure 2 This diagram illustrates the flow steps of the credit risk attribution method disclosed in the embodiments of this specification;

[0025] Figure 3 This diagram illustrates the data processing flow of the decision tree mining module disclosed in the embodiments of this specification;

[0026] Figure 4 This illustration shows the data processing flow of the disassembly module for the support structure and point estimation disclosed in the embodiments of this specification;

[0027] Figure 5 This diagram illustrates the data processing flow of the attention coefficient calculation module disclosed in the embodiments of this specification;

[0028] Figure 6 This diagram illustrates the data processing flow of the strategy contribution calculation module disclosed in the embodiments of this specification;

[0029] Figure 7 This specification illustrates the integration disclosed in the embodiments. Figures 3 to 6 The implementation architecture of the entire attribution analysis process for the corresponding core modules;

[0030] Figure 8 This illustration shows the functional structure of the credit risk attribution device disclosed in the embodiments of this specification. Detailed Implementation

[0031] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0032] As mentioned earlier, attribution analysis for abnormal fluctuations in credit risk indicators mainly relies on manual operation by analysts. This traditional model has significant limitations at the technical level, which restricts the platform's analytical capabilities and response efficiency.

[0033] Specifically, when faced with a full-domain feature set containing hundreds or even thousands of feature dimensions, manual analysis is not only labor-intensive and inefficient, but also makes it difficult to systematically locate the key feature combinations that lead to risk fluctuations from complex feature interactions.

[0034] Based on the above observations and analysis, this specification discloses an automated, data-driven credit risk attribution scheme (hereinafter referred to as the improved scheme or this scheme), which aims to achieve intelligent analysis of the entire process through a computer system.

[0035] Figure 1 The implementation architecture of the improved solution is shown. Under this architecture: First, key feature subsets are automatically selected from the global features using feature dimensionality reduction technology, effectively focusing the analysis scope. Then, a labeled dataset with risk labels is constructed based on the key feature subsets, automatically mining a decision tree path that can clearly characterize high-risk user groups. In this way, the fluctuations in the risk indicators of the overall customer group can be traced back to the target user group defined by the combination of key features, thereby achieving rapid, accurate, and automated root cause localization.

[0036] The specific implementation steps of the above improvement plan will be described in detail below.

[0037] First, it should be noted that the credit risk attribution method disclosed in the improvement plan is executed by a computer system (hereinafter referred to as the system). The computer system can be any hardware device, software platform, independent server or distributed device cluster with data computing and instruction processing capabilities. Its core is to execute the following method steps through programmed instructions, rather than relying on specific physical devices.

[0038] Figure 2 The credit risk attribution method illustrated in the diagram includes the following process steps:

[0039] Step S210: Based on the abnormal fluctuations of credit risk indicators in the credit business, determine the credit user group to be analyzed; Step S220: Perform dimensionality reduction processing on the original feature set involved in the credit user group to obtain a key feature subset; Step S230: Based on the feature values ​​of each user in the credit user group on the key feature subset and the corresponding risk labels, construct a labeled dataset; the risk labels indicate whether the user is in a preset risk state, and the definition of the preset risk state depends on the credit risk indicators; Step S240: Based on the labeled dataset, determine a decision tree path that meets predetermined standards; the decision tree path includes feature items and corresponding value conditions, and the predetermined standards include: the coverage of the decision tree path for users in the labeled dataset who are in the preset risk state reaches a predetermined threshold; Step S250: Based on the decision tree path, generate the attribution conclusion of the abnormal fluctuations.

[0040] The steps above are explained in detail below:

[0041] First, in step S210, the credit user group to be analyzed is determined based on the abnormal fluctuations of credit risk indicators in the credit business.

[0042] It should be understood that credit business and related credit products can include: revolving credit products, such as credit card accounts; fixed-term loan products, such as personal consumer loans and auto loans; scenario-based credit payment services, such as various "buy now, pay later" services; and other forms of credit granting and lending services. Furthermore, in actual credit risk attribution analysis, the analysis typically focuses on a specific credit business or product.

[0043] Credit risk indicators are core metrics used to quantify the asset quality of various credit businesses and products. For example, in revolving and fixed-term credit products such as credit cards and personal consumer loans, indicators such as the First Delinquency Payment Rate (FDP) and the non-performing loan ratio are commonly used to monitor fluctuations in asset quality in real time. Abnormal changes in these indicators directly reflect the level of credit risk faced by the relevant credit products.

[0044] Abnormal fluctuations refer to significant changes in credit risk indicators within a set observation period, compared to their values ​​within a benchmark period, exceeding predetermined thresholds. The benchmark period refers to the historical timeframe selected as the comparison point, whose data typically represents the system's normal or expected risk level. The observation period refers to the current or recent timeframe to be analyzed, whose data needs to be compared with the benchmark period to identify anomalies. For example, in a monthly analysis scenario, the current month (e.g., June 2024) can be set as the observation period, and the previous month (May 2024) or the same period last year (June 2023) can be set as the benchmark period.

[0045] Abnormal fluctuations in credit risk indicators can be automatically triggered through real-time monitoring of the business database, or they can be determined by the system through natural language processing technology by parsing and extracting relevant indicators and periodic information from event description information (or risk attribution questions) input by users (such as risk control analysts).

[0046] The aforementioned group of credit users to be analyzed is typically defined as the set of all or a sample of users who used the credit business or its specific products during the observation period associated with the abnormal fluctuations and the benchmark period used as a comparison benchmark.

[0047] For example, the implementation of this step may include: The system first obtains the time parameters of the observation period and the baseline period associated with the abnormal fluctuation event that triggered the analysis, and simultaneously receives the target credit product identifier issued through the configuration interface or strategy engine as a business screening condition. Based on these parameters, the system extracts a set of user identifiers that have business records within the two periods from the user database, and can perform compliant sampling according to the data volume and computing resources. In this way, it can be ensured that the subsequent analysis object (i.e., the credit user group) is a user behavior dataset that is homogeneous in terms of time range and business attributes, thereby laying a reliable data foundation for generating accurate and interpretable attribution conclusions.

[0048] It should be noted that, for the sake of clarity, the credit risk indicators that require attribution analysis and exhibit abnormal fluctuations may be referred to as abnormal fluctuation indicators, and the credit user group to be analyzed may be referred to as the target analysis group.

[0049] Next, in step S220, the original feature set involved in the target analysis group is subjected to dimensionality reduction processing to obtain a key feature subset.

[0050] It needs to be understood that the original feature set (e.g.) Figure 3 The full-domain feature set shown can be an initial feature set integrated from multiple data sources such as user basic attributes, credit transaction behavior, historical performance records, and device and environmental information. The dimensionality of the original feature set can typically reach hundreds to thousands of dimensions, including features such as the user's occupation type, number of recent applications, number of historical overdue payments, and IP location of the current application.

[0051] Dimensionality reduction aims to filter out redundant features and noise interference in the original feature set, allowing the analysis process to focus more on the core feature dimensions that are truly related to the current risk fluctuations. This improves the accuracy and interpretability of attribution conclusions, while significantly improving the computational efficiency of subsequent modeling and avoiding the impact of the curse of dimensionality on model performance.

[0052] In one embodiment, the original feature set can be preprocessed before dimensionality reduction, for example... Figure 3 As shown in the image:

[0053] 1) Missing value handling refers to the process of identifying and processing missing or unrecorded values ​​in feature data. For example, the process includes: calculating the missing rate of each feature (i.e., the proportion of samples with missing data for that feature to the total number of samples); directly deleting feature items with a missing rate higher than 50%; imputing feature items with a missing rate between 10% and 50% using the median or mode; and using the K-nearest neighbor algorithm for prediction and imputation of feature items with a missing rate lower than 10%.

[0054] 2) Standardization refers to the process of scaling feature data proportionally to bring it within a specific range of mean and standard deviation. For example, this can be achieved using a formula... Processing is carried out, among which The characteristic mean, The standard deviation is denoted as , and the processed feature values ​​conform to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0055] 3) Collinearity detection refers to the process of detecting whether there is a high linear correlation between features. By calculating the correlation coefficient between every two feature terms, redundant feature terms with a correlation coefficient greater than 0.95 are deleted.

[0056] Dimensionality reduction can utilize expert knowledge bases (see [link to relevant documentation]). Figure 3 This is achieved through the dimensionality reduction processing branch 2) shown in the diagram. The expert knowledge base is a structured database that records various historical macro-level events (such as major policy adjustments, natural disasters, etc.) and predefines the specific customer groups that may be affected by each event (such as "service industry practitioners" and "high mortgage customers") and the corresponding risk feature set (such as "industry code", "occupation type", and "personal income stability"). By constructing such a knowledge base, the system can transform abstract external macro-level events into concrete and analyzable feature dimensions.

[0057] In practical applications of the expert knowledge base, the system matches the original feature set (or its pre-processed feature set) with the pre-stored risk feature set in the expert knowledge base. For example, the system iterates through each macro-event entry in the knowledge base, compares its associated risk feature set with the current original feature set, and identifies the intersection between the two. Subsequently, the system uses all successfully matched feature items to construct the key feature subset retained after this dimensionality reduction process. This process effectively integrates the prior knowledge of domain experts into the automated analysis workflow, ensuring that the key feature subset not only has statistical significance but also clear business interpretability.

[0058] Considering the possibility of expert knowledge base matching failures (e.g., the number of features matched by the expert knowledge base is lower than the preset minimum number of features), a fallback approach combining principal component analysis (PCA) is proposed. For more details, please refer to [link to relevant documentation]. Figure 3 The dimensionality reduction processing branch 1 shown in the figure may include the following process steps:

[0059] 1) When feature matching using an expert knowledge base fails, perform PCA analysis on the feature data of each user in the target analysis group under the original feature set to obtain the selected principal components. For example, calculate the covariance matrix of the standardized data of the target analysis group under the original feature set, and obtain a set of uncorrelated principal components through eigenvalue decomposition; then select the top k principal components according to the cumulative variance contribution rate criterion, typically with k ranging from 4 to 8.

[0060] 2) Select the k original features that contribute the most to the principal components to form the key feature subset. For example, this can be achieved by analyzing the loading weights of each original feature on the selected principal components, selecting the top few original features that contribute the most to the principal components, thus forming the key feature subset. It should be understood that "several" in this text refers to one or more in quantity.

[0061] By combining PCA branches, it can be ensured that representative risk features can still be effectively extracted even when matching fails or prior knowledge of the expert knowledge base is lacking.

[0062] It should be noted that Figure 3 The document also illustrates the step of interpreting the meaning of principal components. This step aims to assign explicit business meaning to each principal component (PC) extracted by principal component analysis (PCA), thereby transforming them from abstract mathematical variables into interpretable risk dimensions. Specifically, the interpretation of the meaning of principal components can be performed manually or automatically.

[0063] In the automated interpretation method, the system first automatically filters high-load features under each principal component according to the preset load threshold. Then, it generates the business meaning name and description of the principal component based on the list of high-load features and the background business description by using a preset feature-business topic mapping dictionary or calling a large language model.

[0064] For example, the interpreted business meaning of the principal components may include:

[0065] PC1 (Comprehensive Credit Capability): Defined by high-load features such as "Personal Income" (load 0.85), "Assets" (load 0.82), and "Payment History" (load 0.78), it represents the user's basic credit strength.

[0066] PC2 (Repayment Stability): The key load features are "Repayment Regularity" (0.88) and "Amount Stability" (0.84), while "Frequency of Overdue Payments" (-0.79) is a significant negative load, which together characterize the user's repayment behavior pattern.

[0067] PC3 (Liquidity Stress): Characterized by high load features such as "Debt Ratio" (0.86), "Credit Utilization Rate" (0.83), and "Multiple Borrowings" (0.76), it reflects the degree of financial strain on users.

[0068] PC4 (Environmental Sensitivity): Its fluctuations are positively correlated with "Industry Risk" (0.81) and "Occupational Stability" (0.74), and negatively correlated with "Regional Economy" (-0.77), representing the impact of external environmental factors on users.

[0069] The principal component business meanings explained through the above process can be directly used as the basis for subsequent risk fluctuation attribution analysis, so as to quickly locate risk fluctuations to specific business dimensions, thereby providing key clues and directions for root cause analysis.

[0070] As shown above, the key feature subset can be obtained through dimensionality reduction.

[0071] Then, in step S230, a labeled dataset is constructed based on the feature values ​​of each user in the target analysis group on the key feature subset and the corresponding risk labels.

[0072] It should be noted that the risk label is used to indicate whether an individual user is in a preset risk state. The specific definition of the label can adopt different logics depending on the goal of the attribution analysis. For example, if the goal is to identify new risk events, the risk label can indicate whether an individual user "newly entered a preset risk state during the observation period and was not in this state during the historical period used as a baseline." Or, if the goal is to characterize the overall risk state, the risk label can indicate whether an individual user "was in a preset risk state during the observation period."

[0073] The definition of the aforementioned preset risk status depends on the abnormal fluctuation indicators targeted in this attribution analysis. For example, if the abnormal fluctuation indicator is the first-period delinquency rate, the preset risk status is "first-period delinquency has occurred"; if the abnormal fluctuation indicator is the non-performing loan rate, the preset risk status is "entering a non-performing state". Risk tags transform macro-level business indicator fluctuations into micro-level user-level labels, providing monitoring signals for the subsequent mining of high-risk user profiles.

[0074] For example, if the abnormal fluctuation indicator targeted in this analysis is the first-period delinquency rate, then the risk label can be defined as whether the user has experienced a first-period delinquency within the observation period; if the indicator is the non-performing rate, then it can be defined as whether the user has entered a non-performing state.

[0075] Risk labels can be generated from various sources. For example, they can be automatically generated based on clearly defined business rules and historical performance data, such as overdue status recorded in a post-loan management system. Alternatively, for complex risk scenarios lacking clear rules, large-scale models can be used to intelligently label user behavior data. For instance, a large model can determine whether a user is a "potential fraud risk" user based on their multi-dimensional behavioral sequences.

[0076] In this step, each user's specific values ​​on a subset of key features can be used as a feature vector, and their corresponding risk label can be used as the target variable to form a sample. Samples from all users in the target analysis group are then aggregated to form a labeled dataset.

[0077] Next, in step S240, based on the labeled dataset, a decision tree path that meets predetermined criteria is determined; the decision tree path includes feature terms and corresponding value conditions, and the predetermined criteria include: the coverage rate of the decision tree path for users in the preset risk state reaches a predetermined threshold. It should be understood that users who have entered the preset risk state may be referred to as high-risk users in this document.

[0078] In one implementation, drawing on the advantages of both greedy strategies and decision tree algorithms, a decision tree-like traversal algorithm is proposed. The core idea of ​​this algorithm is to employ a greedy strategy at each decision node, selecting the currently optimal splitting feature, and constructing a decision path through this "locally optimal, step-by-step" approach. The specific implementation process is as follows:

[0079] 1) Based on the labeled dataset, calculate the discrimination ability index of each key feature against the risk label, and select the key feature with the strongest discrimination ability and its corresponding feature value conditions as the first layer node of the decision tree path.

[0080] For example, the ability to distinguish can be evaluated using indicators such as information gain or the Gini coefficient. Assuming that "historical overdue number" has the strongest distinguishing ability, and its optimal classification condition is "≥3 times", then this feature and its condition are used as the first-level node of the decision path.

[0081] 2) In the subset of labeled data defined by the first-level node, calculate the discrimination ability index of the remaining key features against the risk label, and select the key features with the strongest discrimination ability and their corresponding feature value conditions as the second-level node of the decision tree path.

[0082] For example, in the subset of users who meet the first-level condition (i.e., ≥3 historical overdue payments), the discriminative power of the remaining features is further evaluated. Assuming that "number of applications in the last 3 months" has the strongest discriminative power in this subset, and the optimal condition is "≥5 times", then it is designated as a second-level node.

[0083] 3) Recursively execute the above steps until the coverage ratio of the generated decision tree path to high-risk users reaches the predetermined threshold.

[0084] The advantages of this implementation are twofold: firstly, it inherits the interpretability of the decision tree algorithm, with each path clearly expressed as a rule-based structure of "feature A satisfies condition X and feature B satisfies condition Y..."; secondly, it avoids constructing a complete decision tree through a greedy strategy, only requiring layer-by-layer expansion along the optimal path, significantly reducing computational load. For example, compared to constructing a complete decision tree with hundreds of nodes, this implementation may only require a few or a dozen feature evaluations to obtain the target path, significantly improving computational efficiency and making it more suitable for real-time risk analysis scenarios.

[0085] Furthermore, in one embodiment, if the depth of the generated decision tree path has reached a preset depth limit (e.g., 3 layers), but the coverage ratio for high-risk users has not yet reached a predetermined threshold, the system will backtrack to the root node and restart the aforementioned decision tree traversal process from the remaining candidate features to generate a new, different decision tree path. This process is repeated until the cumulative coverage ratio of all generated decision tree paths for high-risk users reaches the predetermined threshold.

[0086] The advantage of this design lies in its robustness and flexibility. By constructing multiple paths to approximate complex risk patterns, it effectively avoids the problem of insufficient coverage caused by the depth limitations of a single path. For example, the first path might identify a typical high-risk customer group with "≥3 historical delinquencies and an average personal income decline of over 30% in the past 3 months," while the second path might discover a potential risk group in first-tier cities with "≥5 applications in the past 3 months." This multi-path collaborative mechanism ensures a comprehensive capture of dispersed and heterogeneous risk groups, thus providing a more complete and robust attribution explanation for abnormal fluctuations while strictly adhering to the constraints of computational complexity (through depth cap control).

[0087] In another implementation, the system first constructs a complete decision tree model based on the labeled dataset, recursively splitting the data until a preset stopping condition (such as node purity, minimum number of samples, etc.) is met, forming a complete decision tree with multiple branches. Subsequently, the system selects paths from all paths from the root node to the leaf node of the decision tree that achieve a predetermined threshold of coverage for high-risk users as decision tree paths.

[0088] As shown above, different implementation methods can be used to obtain (one or more) decision tree paths.

[0089] Then, in step S250, attribution conclusions for the abnormal fluctuations are generated based on the decision tree path.

[0090] In one embodiment, the system can use the selected decision tree path as key input to call a large model for automated attribution analysis. Specifically, the system inputs the combination of feature conditions extracted from the decision tree path (such as "historical delinquency count ≥ 3 times and personal income decline of more than 30% in the past 3 months") and their corresponding risk statistics (such as the delinquency rate of users within the path being 2.5 times higher than the overall rate) into the large model. The large model then combines this with a pre-built business knowledge base to generate an attribution description that conforms to business language conventions, for example: "This risk fluctuation is mainly due to the deterioration of the personal income of existing high-risk customers, especially customers with more than 3 historical delinquencies, whose recent decline in personal income has further impaired their repayment ability."

[0091] In another embodiment, the system can also achieve automated attribution through pre-set rule templates. For example, when the system identifies a combination of features in the decision tree path that includes "occupation type = education and training industry" and "city = Beijing", it can automatically match the "double reduction policy" event in the knowledge base and generate an attribution conclusion: "This risk fluctuation is highly correlated with the adjustment of education industry policies, and the risk performance of education and training practitioners in Beijing has increased significantly."

[0092] On the other hand, in one embodiment, this step can also generate risk control measures (such as raising the credit threshold or tightening credit limits). For example, large models can generate these measures along with their attribution conclusions.

[0093] In summary, the following approach is adopted. Figure 2 The credit risk attribution method presented, or the decision tree path mining module disclosed in the improved solution, first uses feature dimensionality reduction technology to screen out the key features most relevant to risk fluctuations. Then, it uses an automatic mining algorithm to objectively mine feature combination rules that can clearly characterize high-risk groups. This process traces macro-level business indicator fluctuations back to specific user groups defined by micro-level feature combinations. The conclusions are derived automatically based on data and algorithms, thus significantly improving the efficiency, accuracy, and reliability of root cause localization.

[0094] According to another embodiment, the applicant also found that current credit risk attribution methods can only identify which dimensions (such as region, occupation, etc.) are strongly correlated with the increase in risk, for example, concluding that "the risk contribution of the **city customer group has increased significantly".

[0095] However, such methods cannot further reveal the intrinsic driving forces behind the increased risk in this dimension. In fact, the increase in the risk contribution of this dimension may stem from two fundamentally different reasons:

[0096] First, there are changes in customer structure (changes in customer distribution): the proportion (or spending structure) of this customer group (such as the **city customer group) in the overall business has increased significantly. Even if the risk level (which can be understood as quality) of this customer group itself remains unchanged, its expansion will naturally push up the overall risk level.

[0097] Second, changes in customer quality: While the business share of this customer group remains unchanged, the inherent risk level (such as measured by the "first delinquency rate") has deteriorated, or in other words, the customers in this dimension have become worse.

[0098] The risk management strategies for these two reasons are drastically different. If the change is in customer demographics, the focus should be on adjusting the customer structure and controlling the entry or size of high-risk customers. If the change is in customer quality, a thorough investigation and remediation of any vulnerabilities in the risk control strategy for that customer group are necessary. Traditional methods cannot distinguish between these two reasons, making it difficult for business operators to implement precise measures.

[0099] To address the core issue of the inability to distinguish between structural and qualitative changes, this solution proposes a newly designed decomposition module for supporting structure and point estimation. This module decomposes the risk change value (i.e., the risk value during the observation period minus the risk value during the benchmark period, where "risk value" refers to the index value of the abnormal volatility indicator) into structural effects and point estimation effects. Furthermore, by constructing different counterfactual business scenarios and calculating the deviations of these scenarios from historical benchmarks, the independent contributions of these two key effects are quantified. (See [link to relevant documentation]). Figure 4 .

[0100] Specifically, the definitions and calculation formulas for the two key effects are as follows:

[0101] 1) Structural Effect: Used to quantify risk fluctuations caused by changes in customer spending patterns. The calculation logic involves constructing a "counterfactual" scenario—assuming the customer risk level (point estimate) remains at the level of the observation period, but forcibly reverting their spending pattern to the level of the baseline period. The difference between the calculated risk value (baseline period spending pattern × observation period point estimate) and the actual baseline period risk value (such as baseline period FPD) reveals the impact of "more or fewer customers."

[0102] Therefore, the formula for calculating structural effects can be:

[0103] Structural effect = (Base cycle expenditure structure × Observation period point estimate) − Base cycle risk value (1)

[0104] 2) Point Estimation Effect: Used to quantify risk fluctuations caused by changes in the quality of the customer base itself. The calculation logic is to construct another "counterfactual" scenario—assuming the customer base's spending structure remains at the level of the observation period, but forcibly regressing its risk level (point estimate) to the level of the baseline period. The difference between the calculated risk value (observation period spending structure × baseline period point estimate) and the actual baseline period risk value (such as baseline period FPD) reveals the impact of "customers getting better or worse."

[0105] From the above, the formula for calculating the point estimation effect can be:

[0106] Point estimation effect = (Observation period expenditure structure × Benchmark period point estimation) − Benchmark period risk value (2)

[0107] It can be proven that the sum of the structural effect and the point estimation effect is equal to the overall risk change over the observation period and the baseline period, i.e.: Risk value during the observation period − Risk value during the baseline period ≈ Structural effect + Point estimation effect. This mathematical relationship verifies the rigor and completeness of this decomposition method.

[0108] The aforementioned formulas (1) and (2) for calculating the structural effect and the point estimation effect illustrate the newly proposed attribution idea from a macro-level perspective. In order to implement the above calculations, it is necessary to decompose the macro-level "expenditure structure" and "point estimation" to a more granular data level for calculation.

[0109] Specifically, this decomposition is achieved through feature binning technology. After selecting the predetermined feature to be analyzed (such as the number of historical overdue payments), its values ​​are divided into several business-meaning categories, i.e., feature value binning (for example, dividing the number of historical overdue payments into three bins: 0 times, 1-2 times, and ≥3 times). Each bin represents a customer subgroup with relatively consistent internal risk performance.

[0110] Based on this, two core indicators need to be calculated for each sub-container during the baseline period and the observation period:

[0111] a) Structural proportion: This refers to the proportion of the business scale metric (such as number of users or transaction amount) of this sub-group in the overall business scale of the corresponding period. It quantifies the size of this customer subgroup within the whole, thereby quantifying the distribution change, i.e., the structure.

[0112] b) Risk Performance Value: This refers to the proportion of users in this subgroup who triggered the aforementioned credit risk indicators (or, in other words, the number of users added to the numerator of the indicator statistics), relative to the total number of users in this subgroup (e.g., the FPD7 delinquency rate of this subgroup). It quantifies the inherent risk level of this customer subgroup, i.e., its quality.

[0113] Based on these bin-level indicators, the macro-level "spending structure × point estimate" calculation can be implemented as the sum of the products of the structural proportion and risk performance value of all bins. Therefore, the aforementioned macro-level formulas (1) and (2) can be specifically expanded into the following bin-based calculation forms:

[0114] Formula for calculating the structural effect by binning:

[0115] Structural effect contribution value = [Σ(Structural proportion of each sub-box in the baseline period × Risk performance value of each sub-box in the observation period)] - Risk value of the baseline period (3)

[0116] The formula for calculating the binning effect of the point estimation is as follows:

[0117] Point estimate effect contribution value = [Σ(the proportion of each sub-bin structure in the observation period × the risk performance value of each sub-bin in the baseline period)] - the risk value of the baseline period (4)

[0118] Based on the above, we can calculate the structural effect contribution value caused by changes in the user group structure, and the point estimation effect contribution value caused by changes in the user group quality.

[0119] The two types of contribution values ​​calculated above can be used to generate attribution conclusions for the abnormal fluctuation indicators of the target analysis group.

[0120] In one embodiment, the system calculates the weights of the structural effect contribution and the point estimation effect contribution in the total effect to determine the core driving force type of risk volatility. The calculation formula is as follows:

[0121] Structural weight = |Contribution of structural effect| / (|Contribution of structural effect| + |Contribution of point estimation effect|) (5)

[0122] Point estimation weight = |point estimation effect contribution value| / (|structural effect contribution value| + |point estimation effect contribution value|) (6)

[0123] Subsequently, the system can generate corresponding attribution conclusions and treatment suggestions based on preset rules. For example, Figure 4 The decision-making rules shown are as follows: if the structural weight is greater than the predetermined weight threshold (e.g., 70%), it is judged as "structural risk", the conclusion is "deterioration of customer quality", and it is recommended to "optimize the access strategy and control the inflow of high-risk customers"; if the point estimated weight is greater than the predetermined weight threshold, it is judged as "quality risk", the conclusion is "deterioration of homogeneous customer groups", and it is recommended to "strengthen loan monitoring and improve early warning capabilities"; if neither of the weights exceeds the predetermined weight threshold, it is judged as "mixed risk", and it is recommended to "implement both access and loan strategies".

[0124] In another embodiment, the system can use the two types of contribution values ​​obtained from the calculation as key inputs, either individually or together with the aforementioned extracted decision tree path (high-risk user profile) into the large language model. The large language model will automatically generate attribution conclusions and analysis reports that conform to human language habits and have business logic depth.

[0125] On the other hand, after completing the initial attribution of overall risk fluctuations (e.g., determining that the risk is mainly driven by structural effects in the "region" dimension), considering that business stakeholders not only want to know "which dimension" and "which effect" is the main cause, but also want to clarify which specific sub-customer groups under that dimension the risk is concentrated on, so as to invest limited resources in the groups that need the most attention.

[0126] To address this, this solution proposes a recursive drill-down analysis mechanism, aiming to further pinpoint the internal risk focal points based on the identified dominant dimensions. The specific implementation is as follows:

[0127] S41, Identifying the Dominant Feature: The system first calculates the structural effect contribution and point estimation effect contribution of each of several predefined features (e.g., region, occupation, historical delinquency count) to the overall risk change. Then, it aggregates the structural contributions of all features to obtain the total structural change contribution, and aggregates the qualitative contributions of all features to obtain the total qualitative change contribution. By comparing these two totals, it determines whether the current risk fluctuation is primarily driven by "structural change" or "qualitative change," and selects the feature with the largest contribution as the dominant feature (e.g., determining "region" as the dominant feature).

[0128] S42, Target Risk Bins Screening: After determining the dominant feature, the system will conduct in-depth analysis of each bin under that feature. Based on the risk performance value of each bin during the observation period (i.e., the delinquency rate of that bin), it will screen out one or more target bins with the highest risk (for example, under the "Region" feature, it is found that the risk performance value of the "**City" bin is significantly higher than that of other city bins).

[0129] S43, Recursive Drill-Down Analysis: The system then focuses its analysis on the user subgroups defined by the target bins selected in the previous step (e.g., targeting "users in all n-tier cities"). Within this narrowed scope, the system excludes the initial dominant feature (i.e., "region") and then re-executes the aforementioned attribution analysis process on the remaining features (such as occupation, historical delinquency count). This includes recalculating the structural proportions and risk performance values ​​of each bin under these features in the baseline and observation periods, and accordingly calculating their structural effect contribution and point estimate effect contribution at this subgroup level.

[0130] By drilling down recursively (e.g., from "customer group in ** city" to "customer group in ** city with ≥3 historical overdue payments"), the transmission path and accumulation points of risk can be revealed layer by layer, thus providing solid data support for formulating more precise risk control strategies. For example, a strategy can be formulated: tighten the admission criteria for customer group in ** city with ≥3 historical overdue payments.

[0131] In summary, the "support structure + point estimation" decomposition module proposed in this solution for risk changes brings several beneficial effects compared to traditional methods:

[0132] 1. A fundamental improvement in attribution accuracy and decision-making effectiveness

[0133] Traditional methods can only pinpoint the relevant dimensions of risk escalation, while this solution precisely breaks down the overall risk change into structural effects and point estimation effects, achieving a quantitative distinction of the intrinsic driving forces of risk. This allows business stakeholders to clearly identify whether risk fluctuations stem from "changes in customer structure" or "changes in customer quality." Based on this, the system can automatically generate more targeted handling suggestions (e.g., "optimize access" for structural risks, and "strengthen loan monitoring" for quality risks), thereby fundamentally solving the problems of unclear strategy direction and resource misallocation caused by ambiguous attribution, and greatly improving the effectiveness of risk response measures.

[0134] 2. Significantly enhanced depth of risk insight and degree of automation in analysis.

[0135] The proposed recursive drill-down analysis mechanism extends risk attribution from a single feature dimension to a more granular customer segmentation level. It can drill down layer by layer from dominant features (such as "region") to specific risk bins, clearly outlining the transmission path and accumulation points of risks. This automated, progressive analysis capability makes risk identification more precise, allowing risk control resources to be allocated effectively to the highest-risk sub-customer groups, while significantly reducing the workload and time cost of manual multi-dimensional cross-analysis.

[0136] 3. A good balance between the interpretability of technical implementation and its integration with business applications.

[0137] The decomposition analysis of the two effects is based on techniques with clear mathematical logic and business implications, such as counterfactual inference and feature binning. The entire attribution process is transparent, traceable, and verifiable, overcoming the shortcomings of unclear decision-making logic in "black box" models and greatly enhancing the interpretability and credibility of the analysis results.

[0138] According to another example, the applicant also observed that risk managers often face the challenge of limited analytical resources in the daily monitoring of credit risk. Traditional attribution analysis methods tend to focus on those large customer groups (e.g., the 30% of customers in first-tier cities), because their absolute contribution to risk indicators is usually the highest. However, this approach has a blind spot: it easily overlooks those small but rapidly deteriorating "high-risk" dimensions. This interference from the "scale effect" causes many potential risk points to fail to receive timely attention in the early stages, only to be discovered when the risk erupts on a large scale, by which time the best opportunity for intervention has been missed.

[0139] To address the aforementioned issues, this solution also includes a newly designed attention coefficient calculation module. The attention coefficient calculation employs normalization to effectively eliminate the influence of scale factors, enabling analysts to focus on the essential changes in risk. The specific calculation steps are as follows (which can be combined with...). Figure 5 (to understand)

[0140] S51, Data Acquisition. Acquire the following data for the overall credit customer group within the baseline and observation periods:

[0141] 1) Overall risk indicator values ​​(e.g., national total FPD30 delinquency rate);

[0142] 2) The base for calculating overall risk indicators (e.g., the total number of loans issued nationwide).

[0143] Simultaneously, for the target user group (hereinafter referred to as "target customer group", such as "a customer group in a third-tier city") defined by the target feature value binning of the predetermined feature items, we obtain their data for each period in the baseline period and the observation period:

[0144] 1) Risk indicator values ​​(e.g., the FPD30 delinquency rate of the customer group in this city);

[0145] 2) Risk indicator calculation base (e.g., the number of loan disbursements to customers in this city).

[0146] It should be noted that the “risk indicator value” in this step refers to the indicator value of the above-mentioned abnormal fluctuation indicator; while the “risk indicator calculation base” is the uniform denominator used when calculating the indicator.

[0147] S52, based on the acquired data, calculates two key intermediate indicators:

[0148] 1) Total Contribution Percentage: This indicator quantifies the proportion of contribution to overall risk fluctuations resulting from the combined effects of changes in the target customer group's own risk level and size / structure. Its calculation draws on counterfactual thinking, breaking it down into the sum of the risk change effect and the flow change effect. This sum is then divided by the overall risk change value between the observation period and the baseline period to obtain the group's contribution percentage. The relevant calculation formula is as follows:

[0149] Total contribution percentage = [(risk-based contribution to the overall market) + (flow-based contribution to the overall market)] / (overall risk value of the observation period - overall risk value of the benchmark period) (5)

[0150] Risk-converted contribution to the overall market = (Target customer group observation period risk value - Target customer group benchmark period risk value) × (Target customer group observation period base / Overall observation period base) (6)

[0151] Traffic volume converted to overall market contribution = (Target customer group observation period risk value - Overall observation period risk value) × (Target customer group observation period base - Target customer group baseline period base) / Overall observation period base (7)

[0152] Formula (6) measures the impact of changes in the quality of the target customer group on the overall market, while formula (7) measures the impact of changes in the size and structure of the group on the overall market.

[0153] b) Size Share: This refers to the proportion of the target customer group's risk indicator calculation base within the observation period relative to the overall risk indicator calculation base. This indicator reflects the relative size of the target customer group within the overall business. The specific calculation formula is as follows:

[0154] Size percentage = Base for calculating risk indicators during the observation period of the target customer group / Base for calculating risk indicators during the overall observation period (8)

[0155] S53, Calculation of attention coefficient.

[0156] The ratio between the two key intermediate indicators mentioned above is used to obtain the attention coefficient.

[0157] Attention coefficient = Total contribution percentage / Scale percentage (9)

[0158] The aforementioned attention coefficient, through its normalization characteristics, achieves cross-customer group comparability of risk contribution per unit size, thereby accurately locating those hidden risk points that are "small in scale but extremely harmful".

[0159] Based on the calculated attention coefficient, the system can automatically generate attribution conclusions and output suggested priorities for handling the situation. For example, the following decision rules can be configured (where the predetermined upper and lower thresholds can be 2 and 1, respectively):

[0160] Attention coefficient > predetermined upper limit threshold: This indicates a "small-scale, high-contribution" group whose risk and harm are seriously underestimated and require high-priority attention and in-depth analysis.

[0161] The predetermined lower threshold < attention coefficient ≤ predetermined upper threshold: This indicates that the risk contribution is basically matched with the business scale, and normal attention can be maintained.

[0162] Attention coefficient ≤ predetermined lower threshold: indicates a relatively healthy group that is "large in scale but small in contribution", and can be assigned low priority.

[0163] Below is an implementation case of attention coefficient calculation: In a certain month, the national FPD30 increased by 5bp. Traditional attribution analysis would first focus on the first-tier city customer group, which accounted for 30% and contributed 2bp. However, by calculating the attention coefficient, the system discovered a customer group in a third-tier city, accounting for only 3%, with an attention coefficient as high as 4.2. Although its absolute contribution was only 1.5bp, the risk hazard per unit scale was extremely high. Drill-down analysis revealed that the region was experiencing a wave of layoffs in a specific industry. Based on this, the risk control team promptly implemented targeted strategies, effectively preventing the further spread of risk.

[0164] Based on the above, calculating the attention coefficient yields the following beneficial effects: 1) Eliminates the scale effect: Enables a fair comparison of the risk severity of customer groups of different sizes. 2) Precise resource allocation: Provides a scientific priority ranking basis for limited analysis and risk control resources, enabling proactive and early detection of hidden risk points. 3) Enhanced robustness: This method is insensitive to macroeconomic risk fluctuations and changes in the customer base, maintaining stable judgment capabilities across different business scenarios. 4) Strong interpretability: The entire calculation process is based on clear business logic and mathematical formulas, resulting in intuitive and easy-to-understand results that are readily accepted and trusted by business stakeholders.

[0165] According to another example, in credit risk control practice, accurately quantifying the effectiveness of business strategies is key to optimizing decision-making. Currently, it is relatively mature to evaluate the impact of strategies with control groups (e.g., when testing a new interest rate strategy, some users are retained to continue using the original interest rate) through A / B testing. However, the applicant observed that for strategies that are fully implemented without a control group, the actual impact becomes a "black box" due to the lack of comparable benchmarks, making it difficult to accurately quantify.

[0166] To address the aforementioned issues, this solution also proposes a strategy contribution calculation module. First, it should be clarified that in this solution, the target business strategy refers to one or more strategic interventions implemented within the observation period for the target analysis group or a subset thereof, aimed at influencing their credit behavior or risk performance. Such strategies may include, but are not limited to: credit limit adjustment strategies, differentiated pricing (interest rate) strategies, risk access strategies, or marketing outreach strategies. This target business strategy can be a newly implemented strategy within the observation period, or a continuation or adjustment of a historical strategy. In scenarios with multiple parallel strategies, the target business strategy specifically refers to the particular strategy whose independent contribution needs to be quantified.

[0167] The strategy contribution calculation module is used to implement the following processes (which can be combined with...) Figure 6 (to understand)

[0168] Path 1: Possessing accurate calculation strategy contribution in the control group scenario.

[0169] When a pre-defined test group and control group exist for the target business strategy, the system performs precise calculations. The test group is a subset of user samples subjected to the target business strategy during the observation period; the control group is a subset of user samples used for effect comparison during the same observation period that were not subjected to the target business strategy. The system executes the following steps:

[0170] S61, Homogeneity test: Before calculating the contribution of the strategy, the system first verifies whether the test group and the control group are comparable in the baseline period before the strategy is implemented, so as to ensure the accuracy and reliability of the calculation results.

[0171] Specifically, the system performs the test by comparing the difference rates of key business indicators between the test group and the control group. It should be understood that the key business indicators can be selected as needed, for example, they can be calculated as follows:

[0172] Balance difference rate = | (average balance of test group households - average balance of control group households) / average balance of control group households| × 100% (10)

[0173] The balance refers to the outstanding principal balance of the credit account.

[0174] The difference rate of expenditure amount = | (average expenditure amount of the test group - average expenditure amount of the control group) / average expenditure amount of the control group| × 100% (11)

[0175] Furthermore, the system can determine whether the homogeneity test has been passed based on the calculated indicator values. For example, the judgment rule may include: if the difference rate of all key business indicators is less than a predetermined threshold (e.g., 10%), then the homogeneity test is passed and the contribution calculation step is initiated; otherwise, the system outputs a warning message, indicating that there may be a bias in the sample selection and that re-selection is required.

[0176] S63, Calculation of Strategy Contribution: After passing the homogeneity test, the system uses mathematical reduction to quantify the net impact of the target business strategy on credit risk indicators (such as abnormal volatility indicators).

[0177] The core design principle of the calculation is as follows: by removing data from the test group that was affected by the strategy, a virtual market state assuming the strategy was not implemented is reconstructed. This virtual state is then compared with the actual market state, and the difference is the contribution value of the strategy. The general calculation formula can be:

[0178] The contribution of the strategy to the credit risk indicator = Actual market index value - Virtual market index value (12)

[0179] The virtual market index value is calculated by subtracting the changes in the test group caused by the strategy from the actual market data.

[0180] To aid understanding, three calculation examples are provided below, combining three credit risk indicators.

[0181] 1) Taking the credit risk indicator as an example:

[0182] The strategy's contribution to the balance (100 million yuan) = (average balance per user in the test group - average balance per user in the control group) × number of users in the test group / 100,000,000 (13)

[0183] The average balance per user in the test group and the average balance per user in the control group are the average values ​​calculated within each group during the observation period after the strategy was implemented. Specifically, the average balance per user refers to the average of the total outstanding principal balance in the group (test group or control group) divided by the total number of users in the group.

[0184] This formula directly quantifies the change in the overall credit balance resulting from the implementation of the strategy.

[0185] 2) Taking FDP30 as an example, a credit risk indicator:

[0186] The strategy's contribution to FPD30 (bp) = Actual market FPD30 - Virtual market FPD30 (14)

[0187] Wherein, the virtual market FPD30 = [(market FPD numerator - ΔFPD numerator × number of test group users) / (market FPD denominator - ΔFPD denominator × number of test group users)] × 10000. Furthermore, the ΔFPD numerator and ΔFPD denominator are calculated as follows:

[0188] ΔFPD molecule = Average FPD molecules per household in the test group - Average FPD molecules per household in the control group

[0189] ΔFPD denominator = Average FPD denominator of the test group - Average FPD denominator of the control group

[0190] "Average per household" refers to dividing the total numerator or denominator of the group by its total number of users.

[0191] 3) Taking the collection rate as an example, a credit risk indicator:

[0192] The strategy's contribution to the inflow rate (bp) = Actual market inflow rate - Virtual market inflow rate (15)

[0193] Wherein, the virtual market activation rate = [(market activation numerator - Δ activation numerator × number of test group users) / (market activation denominator - Δ activation denominator × number of test group users)] × 10000. Furthermore, the Δ activation numerator and Δ activation denominator in the formula are calculated as follows:

[0194] Δcatalyst input molecules = Average catalyst input molecules per household in the test group - Average catalyst input molecules per household in the control group

[0195] ΔIncrease Denominator = Average Increment Denominator of the Test Group - Average Increment Denominator of the Control Group

[0196] "Average per household" refers to dividing the total numerator or denominator of the group by its total number of users.

[0197] The above describes Path 1, which focuses on the accurate calculation strategy contribution in scenarios with a control group.

[0198] Path 2: Predictive calculation in scenarios without a control group.

[0199] For target business strategies that lack a control group, this solution predicts the effectiveness by matching historical strategies with paths. The specific steps are as follows:

[0200] Step S62, Matching Alternative Control Groups: The system matches historically tested strategies similar to the current target strategy from the historical strategy library, and uses their control groups as alternative control groups. The matching dimensions may include product type, target customer group, and strategy actions.

[0201] For example, suppose the current target strategy is "to promote a new consumer loan product to young white-collar workers in second- and third-tier cities". The system matches a highly similar historical strategy in the historical database: "a similar consumer loan product that was previously promoted to similar cities and customer groups". This historical strategy had a control group at the time, and the system uses the control group of this historical strategy as the "alternative control group" for the current analysis.

[0202] Step S64, Strategy Contribution Prediction Calculation: After obtaining the alternative control group, the system will reuse the process (S61, S63) in Path 1 to calculate the contribution.

[0203] Step S66: Scale Conversion: This step is optional and is activated when the estimated contribution value is a monetary indicator (such as average balance per household, average return per household) rather than a ratio indicator (such as FPD30, collection rate). The system multiplies the average contribution value per household obtained in step S64 by the implementation scale of the target strategy test group to calculate its estimated contribution to the overall market. The conversion formula can be written as:

[0204] Total estimated contribution = Average estimated contribution per user × Number of users in the current test group (16)

[0205] For example, if step S2b calculates that the strategy will contribute an estimated 500 yuan to the average balance per household, and the current test group has 100,000 users, then through scale conversion, the estimated contribution of the strategy to the total balance is 500 yuan / household × 100,000 households = 50,000,000 yuan.

[0206] The above paths one and two address the issue of evaluating the contribution of a single strategy. Based on the calculated strategy contribution, the system can automatically generate attribution analysis and strategy optimization suggestions: First, all parallel strategies are sorted according to their absolute contribution values ​​to identify the key strategies with the greatest impact on risk volatility; then, based on the positive or negative value and magnitude of the contribution, the system judges the strategy's effectiveness and outputs targeted suggestions—strategies with significant positive contributions are deemed effective and their application scale is recommended to be expanded in subsequent risk control; strategies with significant negative contributions are deemed ineffective or defective and intervention or immediate removal is recommended; strategies with near-neutral contribution values ​​are recommended to be included in the continuous monitoring list.

[0207] However, in complex risk control practices, existing technologies cannot clarify the independent contribution of individual strategies among multiple parallel strategies, nor can they identify whether the strategies produce synergistic enhancements or mutual cancellations, making it difficult for risk control teams to determine which strategies are truly effective.

[0208] Furthermore, when it is necessary to evaluate the combined impact of multiple parallel strategies, this scheme uses the marginal contribution method to decompose the contribution:

[0209] 1) Independent contribution: For strategies with a separate control group, the contribution is calculated directly through path one.

[0210] For example, suppose strategy A (interest rate adjustment) and strategy B (credit limit increase) are launched simultaneously. If strategy A has an independent control group, its independent contribution can be directly calculated through path one, for example: "strategy A independently increased FPD30 by 5bp".

[0211] 2) Marginal contribution: For strategies whose independent contributions cannot be directly calculated, their contribution is obtained using the following formula:

[0212] Total Strategy Effect - Σ(Independent Contribution of Other Strategy) - Other Known Effects

[0213] "Other known effects" refers to factors that have been identified and quantified, and that have a systematic impact on business metrics, in addition to the multiple parallel strategies being evaluated. These factors are typically measured independently using methods such as time series models, external data regression, or historical benchmark analysis. Examples of common factors are as follows:

[0214] ① Seasonal effect: For example, the year-end peak consumption season usually leads to a natural increase in spending.

[0215] ② Holiday effect: For example, increased travel spending during long holidays may have an inherent impact on the delinquency rate of specific products.

[0216] ③ Known macroeconomic fluctuations: For example, the general impact of a uniform interest rate adjustment policy on the overall customer base.

[0217] ④ Natural drift of known traffic or customer structure: changes in customer groups based on long-term trend predictions.

[0218] For example, continuing from the previous example, if the total effect of the strategy is to increase FPD30 by 12bp, the independent contribution of strategy A is 5bp, and the model calculates that FPD30 naturally increases by 1bp due to the "seasonal effect." Then the marginal contribution of strategy B is 12bp - 5bp - 1bp = 6bp. By deducting the seasonal effect, the evaluation of the effect of strategy B is more accurate.

[0219] 3) Interaction effect analysis:

[0220] Interaction effect = Total policy effect - Σ(Contribution of all individual policies) - Other known effects

[0221] It should be noted that if the obtained interaction effect value exceeds a predetermined threshold (e.g., 10% of the total effect), the system will mark it, indicating that there is a significant synergistic or counteracting effect between the strategies.

[0222] For example, if the independent contribution of strategy A is 5bp, the marginal contribution of strategy B is 6bp, and the deducted "other known effects" is 1bp, but the total effect of both being implemented simultaneously is 15bp, then the interaction effect is 15bp - (5bp + 6bp) - 1bp = 3bp. This value is positive and exceeds the threshold, and the system will indicate that there is a synergistic enhancement effect between strategies A and B.

[0223] Therefore, in the strategy contribution calculation module of this solution, a precise calculation framework for strategy contribution is proposed, integrating homogeneity testing and mathematical reduction methods. An innovative historical strategy matching path is designed to solve the problem of quantifying the effects of strategies without a control group. Furthermore, a marginal contribution method is introduced to effectively decompose independent contributions and interaction effects in multi-strategy scenarios. Thus, a reliable quantitative basis is provided for the full lifecycle management of risk control strategies—from design to implementation to evaluation and iteration—significantly improving the refinement of strategy management and scientific decision-making capabilities.

[0224] The above details the four core analysis modules in this solution: decision tree path mining module, support structure and point estimation decomposition module, attention coefficient calculation module, and strategy contribution calculation module. Examples can be found by referring to the relevant documentation. Figure 3 , Figure 4 , Figure 5 and Figure 6 .

[0225] This solution further proposes that the four core analysis modules mentioned above can be integrated to construct a collaborative and automated attribution system. Each module is connected via a unified scheduling engine and data interface, thereby achieving a closed-loop analysis across the entire chain, from risk signal perception and multi-dimensional attribution localization to strategy effectiveness quantification.

[0226] To visually demonstrate the workflow of this system, we will now combine... Figure 7 The architecture example shown illustrates this. The process begins with responding to user attribution requests. These requests are typically submitted in natural language or structured forms, with the core objective of pinpointing the root cause of fluctuations in specific credit risk metrics. For example, a user might ask: "Analyze the reasons for the significant increase in the 'First Delinquency Rate (FPD30)' among 'young white-collar customers' in the past month," or "Identify the key characteristic dimensions that led to the deterioration of the 'Non-Performing Loan Rate' in the third quarter."

[0227] Step S1: Key information analysis.

[0228] The system first analyzes the risk attribution questions input by the user, extracting the core elements to lay the foundation for subsequent targeted analysis. These core elements may include:

[0229] 1) Credit Risk Indicators: These are quantitative risk measures that indicate abnormal fluctuations and require close monitoring. Examples include the first delinquency rate (FPD30), the non-performing loan ratio (NPL), or the collection rate. The system will use these indicators as the assessment benchmark and calculation object.

[0230] 2) Attribution Type: This refers to the specific direction of attribution analysis required for abnormal risk indicator fluctuations. For example, the analysis objective could be to pinpoint the root cause of the fluctuation, quantify the impact of a specific business strategy on the fluctuation (strategy attribution), or identify the core high-risk customer group causing the fluctuation (event attribution). Clearly defining the attribution type determines the priority and focus of module calls in the subsequent analysis process.

[0231] 3) Observation period: refers to the specific time period during which the risk indicator shows an abnormal state. It is the data time range for analysis, such as "June 2024" or "third quarter".

[0232] For example, when a user inputs the risk attribution question as: "Quantitatively assess the impact of the 'automatic credit limit increase' strategy launched for micro and small business owners in the second quarter on the 'non-performing loan ratio'", the system will parse the following: the credit risk indicator is "non-performing loan ratio", the attribution type is "strategy attribution", and the observation period is "second quarter".

[0233] Step S2: Risk situation assessment.

[0234] Based on the information extracted in step 1, the system determines whether the credit risk indicators have truly deteriorated. If the determination is "yes," then step 3 is executed for in-depth attribution analysis; if the determination is "no," then the process proceeds to step 4.

[0235] The assessment process can combine the following two analytical methods:

[0236] 1) Longitudinal Comparison (Trend Analysis): The system compares the performance of indicators during the observation period with their own historical performance during the same period or multiple consecutive historical periods. The system not only checks for statistically significant differences but also analyzes whether there is a continuous monotonic deterioration trend. For example, by pulling monthly data from the past year, if the target indicator shows an increase for three consecutive months and the cumulative increase exceeds a predetermined threshold, a longitudinal warning is triggered.

[0237] 2) Horizontal Comparison (Benchmark Analysis): The risk performance of the target analysis group during the observation period is compared with other similar customer groups that have similar characteristics but are not affected by suspected factors (as a benchmark for horizontal comparison). For example, comparing the same risk indicator of customer groups from different credit product lines or different locations within the same period. If the risk performance of the target analysis group is significantly worse than that of one or more similar customer groups, and the difference exceeds a predetermined threshold, a horizontal warning is triggered.

[0238] Furthermore, the system can comprehensively judge the results of the two analyses based on predetermined rules. Typically, if either the longitudinal or horizontal comparison triggers an alert, the risk situation is judged to have "deteriorated" (yes), and the system will execute step 3 to initiate in-depth attribution analysis. If neither dimension's comparison triggers an alert, the risk situation is judged to have "not deteriorated" (no), and the system will jump to step 4 to directly generate a conclusion and report. This dual verification mechanism effectively avoids misjudgments caused by normal fluctuations in a single dimension, ensuring the reliability of the assessment results.

[0239] Step 3: Multidimensional deep attribution analysis.

[0240] Once step 2 confirms the deterioration of risk indicators, the system initiates a multi-dimensional in-depth analysis mechanism. This mechanism includes four flexibly invoked analysis paths (corresponding to steps S3.1 to S3.4 below). The system can select to execute one or more of these paths based on the attribution type, data availability, and resource conditions. Under resource-allowed conditions, these analysis paths can also be executed in parallel to improve overall attribution efficiency.

[0241] Step 3.1: Channel Analysis. This step aims to pinpoint the specific sources of risk spillover by breaking down channels from multiple dimensions.

[0242] The system segments the target customer group according to preset channel dimensions (such as funding institutions, customer acquisition platforms, traffic channels, etc.) and calculates the contribution of each sub-channel to the overall risk deterioration, thereby identifying specific channels whose risk performance deviates significantly from the benchmark.

[0243] For example, suppose step 2 determines that the overall FPD30 of the "young white-collar customer group" has deteriorated. The system performs channel analysis, first splitting this customer group into three sub-channels: "Platform A," "Platform B," and "Platform C," based on the "customer acquisition platform." Calculations show that the overall FPD30 deterioration is +15 bp; "Platform A" contributes +3 bp, "Platform B" contributes +2 bp, while "Platform C" contributes a significant +10 bp, and its risk level is significantly higher than other channels and the overall average. The analysis indicates that "Platform C" is the main source of risk for this round of FPD30 deterioration. This identification will guide subsequent analysis resources to focus on the customer group introduced by "Platform C," investigating whether there are specific approval strategy loopholes, fraud concentration, or operational quality issues with channel partners.

[0244] Step 3.2: Risk control process analysis. This step aims to preliminarily identify the main links in risk deterioration by analyzing the linkage between risk indicators in different observation windows, thereby focusing the investigation on specific stages such as "loan strategy approval" or "post-loan management".

[0245] The analysis logic in this step may include:

[0246] 1) If the risk is mainly caused by problems with the loan approval strategy (such as relaxed access rules or overly aggressive credit granting), it usually leads to a deterioration in asset quality from the early stages. This is manifested by a significant increase in early risk indicators (such as FPD4) and mid-term risk indicators (such as FPD30) in the same direction.

[0247] 2) If the risk is mainly caused by post-loan management issues (such as decreased collection efficiency or ineffective outreach strategies), its impact usually manifests at a later stage of the account lifecycle. This is characterized by a significant deterioration in mid-term risk indicators (such as FPD30), while early-stage risk indicators (such as FPD4) remain stable or fluctuate only slightly.

[0248] For example, suppose step 2 determines that the overall non-performing loan ratio of the "Micro and Small Enterprise Loan" product has deteriorated. The system performs risk control analysis, comparing the changes in FPD4 and FPD30 relative to the baseline period:

[0249] Scenario 1 (Suspected Loan Issue): Data shows that FPD4 rose from 0.5% in the baseline period to 0.9% in the observation period, while FPD30 rose from 1.2% to 2.1%. Both increased by more than 50%, and their trends were highly synchronized.

[0250] Conclusion: The risks were already significantly exposed early in the account lifecycle, and the problems likely stemmed from changes in loan approval strategies or biases in customer selection.

[0251] Scenario 2 (Suspected post-loan issues): Data shows that FPD4 remained stable at 0.5% during both the benchmark and observation periods, but FPD30 rose from 1.2% to 1.8%.

[0252] Conclusion: Early risk was stable, but the proportion of accounts that migrated to a worse state after the first overdue payment increased, which may be due to a decline in the efficiency of post-loan collection, customer care and other management processes.

[0253] This analysis provides the system with clear guidance for subsequent in-depth investigations (see steps 3.3 or 3.4 below).

[0254] Step 3.3: Refined Risk Attribution in Loan Processing. This step, as a core analytical step, integrates multiple calculation modules to perform a comprehensive risk attribution of the loan processing stages. The analysis includes both quantifiable and non-quantifiable calculations.

[0255] A. Quantifiable Component: The system can call four modules in parallel to execute the following steps, quantifying risk sources from different dimensions:

[0256] Step 3.3.1: Invoke the decision tree path mining module to automatically identify user feature combination rules that lead to high risk. For details, please refer to the relevant descriptions in the preceding embodiments, for example... Figure 3 This will not be elaborated upon here.

[0257] Step 3.3.2: Invoke the decomposition module for the support structure and point estimation to quantify the respective contributions of changes in customer structure and deterioration in customer quality during risk changes. For details, please refer to the relevant descriptions in the preceding embodiments, for example... Figure 4 This will not be elaborated upon here.

[0258] Step 3.3.3: Invoke the attention coefficient calculation module to identify hidden high-risk customer groups that, although small in scale, are experiencing rapidly escalating risks. For details, please refer to the relevant descriptions in the preceding embodiments, for example... Figure 5 This will not be elaborated upon here.

[0259] Step 3.3.4: Invoke the strategy contribution calculation module to assess the actual impact of recently launched business strategies on risk indicators. For details, please refer to the relevant descriptions in the preceding embodiments, such as... Figure 6 This will not be elaborated upon here.

[0260] B. Non-quantifiable portion (residual analysis):

[0261] Step 3.3.5: Perform a systematic analysis on the rising residuals of the point estimates that cannot be explained by the above model.

[0262] The residual is the difference between the actual observed value of a risk indicator and the portion explained by the aforementioned quantifiable modules (steps 3.3.1 to 3.3.4). It represents the risk information that was not captured at the time.

[0263] Significant residuals indicate the existence of unknown risk factors beyond the coverage of known modules, or suggest the existence of deep, complex causal mechanisms not represented by the model. Tracking these residuals is key to discovering new risk patterns and driving model iteration.

[0264] The system can calculate the magnitude and statistical significance of residuals. When the residuals exceed a predetermined threshold, it will automatically extract multi-dimensional information such as feature data and time series patterns from high residual sample groups. Combined with auxiliary information such as external economic indicators and industry dynamics, it will perform pattern recognition and correlation analysis through an integrated large model analysis module to detect early signals of potential unknown risk factors or new risk patterns.

[0265] Example: Suppose the FPD30 observation value of a certain product increases by 20 bp, and the total contribution calculated by the quantifiable module is 16 bp, then the residual is 4 bp. This residual will trigger an automated analysis process to explore possible unknown risk causes through multi-source information fusion analysis.

[0266] Step 3.4: Post-loan analysis. This step aims to quantitatively assess the effectiveness of each stage of the collection process, identify potential bottlenecks in post-loan management, and determine whether they contribute to risk deterioration.

[0267] The analytical logic for this step may include: the system constructing a complete post-loan outreach funnel, starting from the "pending customer group," and analyzing the conversion from "outreach success rate," "negotiation success rate," to "final repayment rate" layer by layer. By comparing the changes in indicators at each stage during the risk deterioration cycle with those during the baseline cycle, the specific stages where efficiency declines can be identified. Simultaneously, the system will evaluate the performance differences of different collection strategies (such as different outreach channels, outreach timing, and script templates) on the current customer group.

[0268] For example, assuming step 2 determines that the FPD30 of the "young white-collar customer group" has deteriorated, the system performs post-loan analysis, focusing on the collection performance of this customer group during the observation period:

[0269] Scenario 1 (Outreach Failure): Data shows that compared to the baseline period, the first-day SMS reach rate for this customer group decreased from 95% to 88% during the observation period, and the intelligent outbound call connection rate decreased from 75% to 68%. This indicates that a large number of overdue accounts were not effectively contacted in the early stages, missing better collection opportunities, which may lead to overdue accounts migrating to a worse state, thus pushing up FPD30.

[0270] Scenario 2 (Negotiation Phase Failure): Data shows that while the reach rate remains stable, the repayment commitment rate after reaching customers drops from 40% in the baseline period to 30%, and the fulfillment rate after commitment drops from 70% to 60%. This suggests that although customers were contacted, there may be problems in the negotiation phase of collection, or a systemic decline in the repayment willingness / ability of the current customer base, leading to reduced collection efficiency.

[0271] Scenario 3 (Strategy Matching Failure): Analysis of strategy contribution revealed that a newly launched "mild reminder" strategy performed normally across the overall customer base. However, within the specific sub-group of "young white-collar workers," its recovery rate was significantly lower than the traditional "serious reminder" strategy. This indicates that the post-loan strategy failed to achieve refined customer segmentation, leading to resource misallocation and efficiency loss.

[0272] Through the above analysis, the system can clearly identify whether post-loan management is a "source of problems" or "performing normally." If a significant deterioration in key post-loan indicators is found, it is identified as a major cause of increased risk, and targeted optimization suggestions are provided, such as repairing contact channels, optimizing negotiation techniques, or implementing more refined collection strategies. If post-loan indicators remain stable or improve, their dominant role can be excluded from the attribution report, and the focus of analysis can shift to the pre-loan and loan processing stages.

[0273] Step 4: Comprehensive Analysis and Output. Based on the outputs of each analysis path in Step 3, the system conducts a comprehensive analysis using preset analysis rules, a large-scale model analysis module, or a combination of both, to generate final attribution conclusions and action recommendations. This step may include the following three branches:

[0274] Step 4.1: Root cause determination.

[0275] The system integrates the outputs of each module in step 3 and makes judgments through one or more of the following methods: 1) logical inference based on preset causal association rules; 2) calling the large model to perform in-depth semantic analysis and logical reasoning on multidimensional conclusions; 3) combining the rule engine and the large model for collaborative judgment.

[0276] For example, when channel analysis (step 3.1) identifies high-risk channels, decision tree mining (step 3.3.1) identifies specific feature combinations, and strategy contribution calculation (step 3.3.4) confirms the impact of relevant strategies, the system can generate accurate judgments such as "the root cause of the current risk deterioration is that the customer group with feature X introduced by channel C has generated an unexpected negative reaction to the recently launched strategy Y" through rule matching or large model analysis.

[0277] Step 4.2: Trend prediction.

[0278] The system can combine current attribution conclusions with historical risk patterns and make predictions through one or more of the following methods: 1) quantitative extrapolation based on time series forecasting models; 2) trend extrapolation by calling large models and combining business knowledge; 3) hybrid prediction by integrating statistical models and large models.

[0279] For example, if the root cause is determined to be "continuous traffic generation from a high-risk channel", and the traffic from that channel is expected to remain stable, the system can predict, through traditional models or large-scale model deduction, that "this risk indicator will remain at a high level in the next month, with an expected increase of 5-8bp".

[0280] Step 4.3: Strategy Recommendations.

[0281] Based on root causes and trend predictions, the system can output suggestions in the following ways: 1) matching optimization solutions from a preset strategy knowledge base; 2) calling a large model to generate innovative solutions; 3) combining the knowledge base and the large model to output comprehensive suggestions.

[0282] For example, in response to the above root causes and predictions, the system may generate the following recommendations: 1) Immediate measures: Suspend the acquisition of customer groups with characteristic X from channel C; 2) Strategy optimization: Re-evaluate and adjust the application parameters of strategy Y on this customer group; 3) Long-term mechanism: Establish a forward-looking risk assessment framework for channel-introduced customer groups.

[0283] Based on the above, a comprehensive assessment can be made.

[0284] Step 5: Report Generation. This step aims to transform the complete output of the aforementioned analysis process into a highly readable attribution analysis document that can be directly used by business personnel.

[0285] Generally, the system invokes the automated report generation module, which integrates all key outputs from steps 1 to 4 based on a preset report template. To improve the logical coherence and natural language of the report, the system can further invoke the large model text generation module to intelligently summarize the structured analysis results, provide natural language descriptions of causal relationships, and offer personalized suggestions.

[0286] The final "Credit Risk Attribution Analysis Report" can include the following core sections: 1) Executive Summary: A high-level summary of the risk issue, root causes, and core recommendations. 2) Problem Background: Reproduces the user-input attribution request and the key elements analyzed in Step 1. 3) Analysis Process: Clearly demonstrates the key findings and data support from risk situation assessment (Step 2) to multi-dimensional in-depth attribution (Step 3). 4) Comprehensive Conclusion: Systematically presents the root cause determination (Step 4.1) and trend prediction (Step 4.2). 5) Strategy Recommendations: Detailed lists of targeted optimization action plans (Step 4.3). 6) Appendix: May include important data charts, brief explanations of model methods, and in-depth information such as residual analysis.

[0287] In summary, the following approach is adopted. Figure 7 The attribution framework shown can produce the following beneficial effects:

[0288] 1) Full-process automation and efficiency improvement: The solution connects the entire analysis chain from risk perception and multi-dimensional attribution to strategy recommendations, reducing the manual analysis that used to take several days to hours, and greatly improving risk response efficiency.

[0289] 2) Multi-dimensional collaborative positioning for accurate attribution: Through collaborative analysis of channel analysis, risk control link positioning, refined attribution during the loan process and post-loan assessment, the source of risk is accurately identified, avoiding the one-sidedness of a single perspective and ensuring the comprehensiveness and accuracy of the attribution conclusions.

[0290] 3) Intelligent analysis and closed-loop management: Integrating rule engine and big model analysis capabilities, it transforms complex analysis results into clear root causes, trend predictions and specific strategy suggestions, forming a complete management closed loop of "analysis-decision-action".

[0291] This solution addresses the pain points of traditional risk analysis, such as fragmented processes, low efficiency, and reliance on expert experience, providing financial institutions with an efficient, accurate, and practical risk attribution solution.

[0292] Corresponding to the credit risk attribution method described above, embodiments of this specification also disclose a credit risk attribution device, which can be integrated into a computer system. See also Figure 8 The credit risk attribution device 800 may include the following functional units:

[0293] The analysis group determination unit 810 is configured to determine the credit user group to be analyzed based on the abnormal fluctuations of credit risk indicators in the credit business. The dimensionality reduction processing unit 820 is configured to perform dimensionality reduction processing on the original feature set involved in the credit user group to obtain a key feature subset. The labeled data construction unit 830 is configured to construct a labeled dataset based on the feature values ​​of each user in the credit user group on the key feature subset and the corresponding risk labels; the risk labels indicate whether the user is in a preset risk state, and the definition of the preset risk state depends on the credit risk indicators. The path determination unit 840 is configured to determine a decision tree path that meets predetermined criteria based on the labeled dataset; the decision tree path includes feature items and corresponding value conditions, and the predetermined criteria include: the coverage rate of the decision tree path for users in the preset risk state reaches a predetermined threshold. The conclusion generation unit 850 is configured to generate an attribution conclusion for the abnormal fluctuations based on the decision tree path.

[0294] In one embodiment, the dimensionality reduction processing unit 820 is specifically configured to: match the original feature set with a risk feature set in an expert knowledge base; the risk feature set corresponds to the credit business customer groups affected by historical macroeconomic events. If the match is successful, the matched features are used to construct the key feature subset.

[0295] Furthermore, in a specific embodiment, the dimensionality reduction processing unit 820 is also configured to: in the case of matching failure, perform principal component analysis on the feature data of each user in the credit user group under the original feature set to obtain the selected principal components; select the original features that contribute the most to the principal components to form the key feature subset.

[0296] In one embodiment, the path determination unit 840 is specifically configured to: calculate the discriminative power index of each key feature against the risk label based on the labeled dataset, thereby selecting the key feature with the strongest discriminative power and its corresponding feature value conditions as the first-level node of the decision tree path. Within the labeled data subset defined by the first-level node, calculate the discriminative power index of the remaining key features against the risk label, thereby selecting the key feature with the strongest discriminative power and its corresponding feature value conditions as the second-level node of the decision tree path. The above steps are recursively executed until the generated decision tree path conforms to the predetermined standard.

[0297] Furthermore, in a specific embodiment, the path determination unit 840 is further configured to: if the depth of the generated decision tree path has reached a preset depth limit, but still has not reached the predetermined standard, generate one or more additional decision tree paths until the predetermined standard is reached.

[0298] In one embodiment, the abnormal fluctuations involve the observation period and the baseline period for the credit risk indicator; the credit risk attribution device 800 further includes an effect decomposition unit 860, configured as follows:

[0299] For each feature value bin of the predetermined characteristic items of the credit user group, the structural proportion and risk performance value of the bin are calculated in each period of the baseline period and the observation period. The structural proportion refers to the proportion of the business scale of the bin in the total business scale within the corresponding period, and the risk performance value refers to the proportion of the number of users in the bin who triggered the credit risk indicator to the total number of users in the bin. Based on the calculated structural proportion and risk performance value, the structural effect contribution value caused by changes in the user group structure and the point estimation effect contribution value caused by changes in the user group quality are further calculated. The structural effect contribution value and the point estimation effect contribution value are used to generate the attribution conclusion for the abnormal fluctuations.

[0300] Furthermore, in a specific embodiment, there are multiple predetermined feature items; the effect decomposition unit 860 is further configured to: determine the total contribution value of structural change based on the structural effect contribution value of each predetermined feature item; and determine the total contribution value of quality change based on the point estimation effect contribution value of each predetermined feature item. The total contribution value of structural change is compared with the total contribution value of quality change, and a dominant feature item is determined from the multiple predetermined feature items based on the comparison result. Target bins are selected from the dominant feature items based on the risk performance values ​​of each feature value bin under the dominant feature item. The user group defined by the target bins is taken as the object to be analyzed. From the remaining feature items after excluding the dominant feature item, the structural proportion and risk performance values ​​of each feature value bin under each predetermined feature item are re-obtained in the baseline period and observation period, and the structural effect contribution value and point estimation effect contribution value are calculated accordingly for recursive analysis.

[0301] In one embodiment, the abnormal fluctuations involve the observation period and benchmark period for the credit risk indicator; the credit risk attribution device 800 further includes a attention coefficient calculation unit 870, configured as follows:

[0302] The overall risk index values ​​and the overall risk index calculation base are obtained for each period within the baseline and observation periods. For the target user group defined by target feature value binning of predetermined feature items, their risk index values ​​and risk index calculation bases are obtained for each period. Based on the obtained data, the total contribution ratio and size ratio of the target user group are calculated. The total contribution ratio quantifies the degree of contribution to the abnormal fluctuations caused by changes in the group's own risk level and size structure. The size ratio is the proportion of the group's risk index calculation base to the overall risk index calculation base during the observation period. The ratio between the total contribution ratio and the size ratio is calculated as the attention coefficient of the target user group. This attention coefficient is used to generate attribution conclusions for the abnormal fluctuations.

[0303] In one embodiment, the abnormal fluctuations involve an observation period and a baseline period for the credit risk indicator, during which a target business strategy is implemented; the credit risk attribution device 800 further includes a strategy contribution calculation unit 880, configured as follows:

[0304] A test group and a control group corresponding to the target business strategy are identified. The test group is the user subgroup subjected to the target business strategy during the observation period, and the control group is the user subgroup not subjected to the target business strategy during the same period. Based on data from the test group and control group within a baseline period before the implementation of the target business strategy, a homogeneity test is performed on the test group and control group. If the homogeneity test passes, based on data from the test group and control group within the observation period after the implementation of the target business strategy, the true contribution value of the target business strategy to business indicators is calculated; wherein, the true contribution value is used to generate the attribution conclusion for the abnormal fluctuations.

[0305] Furthermore, in a specific embodiment, the strategy contribution calculation unit 880 is further configured to: in the absence of a control group corresponding to the target business strategy, match similar historical strategies of the target business strategy from the historical strategy library, and determine the control group corresponding to the similar historical strategy as a substitute control group. Based on the data of the test group during the observation period after the implementation of the target business strategy, and the data of the substitute control group during the historical observation period after the implementation of the similar historical strategy, calculate the estimated contribution value of the target business strategy to the business indicators. The estimated contribution value is used to generate the attribution conclusion for the abnormal fluctuations.

[0306] It should be noted that for a description of the above functional units, please refer to the relevant description of the process method in the foregoing embodiments.

[0307] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform... Figure 2 The method described.

[0308] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement... Figure 2 The method described.

[0309] In this specification, "Large Language Model" (LLM) is an abbreviation for Large Language Model. A LLM is a natural language processing model based on deep learning techniques, typically with billions to hundreds of billions or even more parameters, possessing powerful language understanding and generation capabilities. LLMs can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing exceptionally well in natural language tasks. LLMs learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them excellent generalization capabilities. The core capabilities of LLMs include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Their usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the LLM to generate specific outputs by designing prompts. Cue words can be task descriptions or instructions in text form, used to stimulate the semantic understanding and generation capabilities of large language models. In fine-tuning mode, large language models are further trained on small-scale datasets in specific domains to optimize their performance on specific tasks. The powerful generalization ability and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0310] In some embodiments, large language models can also understand and generate data from other modalities (such as visual and audio data). In this case, large language models can also be called multimodal large language models (MLLMs). MLLMs provide a richer and more natural interactive experience by integrating multiple types of input and output, such as text, images, and sound. The core advantage of MLLMs lies in their ability to process and understand information from different modalities and fuse this information to complete complex tasks. For example, MLLMs can analyze an image and generate descriptive text, or generate a corresponding image based on a text description. This cross-modal understanding and generation capability makes MLLMs widely applicable across multiple fields.

[0311] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025), and will not be repeated here.

[0312] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0313] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A credit risk attribution method, executed by a computer system, comprising: Based on abnormal fluctuations in credit risk indicators in credit business, identify the credit user group to be analyzed; The original feature set involving the credit user group is subjected to dimensionality reduction processing to obtain a key feature subset; Based on the feature values ​​of each user in the credit user group on the key feature subset and the corresponding risk labels, a labeled dataset is constructed; the risk labels indicate whether the user is in a preset risk state, and the definition of the preset risk state depends on the credit risk indicators. Based on the labeled dataset, determine the decision tree path that meets the predetermined criteria; The decision tree path includes feature terms and corresponding value conditions, and the predetermined standard includes: the coverage of the decision tree path for users in the predetermined risk state reaches a predetermined threshold. Based on the decision tree path, the attribution conclusions for the abnormal fluctuations are generated.

2. The method according to claim 1, wherein, The dimensionality reduction process includes: The original feature set is used to match a risk feature set in an expert knowledge base; the risk feature set corresponds to the credit business customer groups affected by historical macroeconomic events. If a match is successful, the key feature subset is constructed using the matched features.

3. The method according to claim 2, wherein, The dimensionality reduction process also includes: In the event of a failed match, principal component analysis is performed on the feature data of each user in the credit user group under the original feature set to obtain the selected principal components. The top k original features that contribute the most to the principal components are selected to form the key feature subset.

4. The method according to claim 1, wherein, Based on the labeled dataset, determine decision tree paths that meet predetermined criteria, including: Based on the labeled dataset, the discrimination ability index of each key feature against the risk label is calculated, and the key feature with the strongest discrimination ability and its corresponding feature value conditions are selected as the first layer node of the decision tree path. In the subset of labeled data defined by the first-level node, the discrimination ability index of the remaining key features against the risk label is calculated, and the key features with the strongest discrimination ability and their corresponding feature value conditions are selected as the second-level node of the decision tree path. The above steps are executed recursively until the generated decision tree path meets the predetermined criteria.

5. The method of claim 4, further comprising: If the depth of the generated decision tree path has reached the preset depth limit, but still has not reached the predetermined standard, then one or more additional decision tree paths are generated until the predetermined standard is reached.

6. The method according to claim 1, wherein, The abnormal fluctuations involve the observation period and benchmark period for the credit risk indicators; the method further includes: For each characteristic value bin of the predetermined characteristic items of the credit user group, calculate the structural proportion and risk performance value of the bin in each period of the baseline period and the observation period; wherein, the structural proportion refers to the proportion of the business scale of the bin in the total business scale in the corresponding period, and the risk performance value refers to the proportion of the number of users in the bin who trigger the credit risk indicator to the total number of users in the bin. Based on the calculated structural proportion and risk performance value, the structural effect contribution value caused by changes in user group structure and the point estimation effect contribution value caused by changes in user group quality are further calculated. The structural effect contribution value and the point estimation effect contribution value are used to generate the attribution conclusions for the abnormal fluctuations.

7. The method according to claim 6, wherein, The predetermined feature items are multiple; the method further includes: Based on the structural effect contribution values ​​of each predetermined feature item, the total contribution value of structural change is determined; and based on the point estimation effect contribution values ​​of each predetermined feature item, the total contribution value of quality change is determined. The total contribution value of the structural change is compared with the total contribution value of the mass change, and the dominant feature is determined from the plurality of predetermined feature items based on the comparison result; Based on the risk performance value of each feature value bin under the dominant feature item, target bins are selected from the dominant feature item; The user group defined by the target bin is taken as the object to be analyzed. Among the remaining features after excluding the dominant feature items, the structural proportion and risk performance value of each feature value bin under each predetermined feature item are re-acquired in the baseline period and observation period. Based on this, the structural effect contribution value and the point estimation effect contribution value are calculated for recursive analysis.

8. The method according to claim 1, wherein, The abnormal fluctuations involve the observation period and benchmark period for the credit risk indicators; the method further includes: Obtain the overall risk index value and the overall risk index calculation base for each period in the baseline period and observation period, and for the target user group defined by the target feature value binning of the predetermined feature items, obtain the risk index value and risk index calculation base for each period. Based on the acquired data, the total contribution ratio and the size ratio of the target user group are calculated; wherein, the total contribution ratio quantifies the degree of contribution of the group's own risk level changes and size structure changes to the abnormal fluctuations, and the size ratio is the proportion of the group's risk indicator calculation base to the overall risk indicator calculation base during the observation period; The ratio between the total contribution percentage and the scale percentage is calculated as the attention coefficient of the target user group; wherein, the attention coefficient is used to generate the attribution conclusion of the abnormal fluctuation.

9. The method according to claim 1, wherein, The abnormal fluctuations involve an observation period and a benchmark period for the credit risk indicators, during which the target business strategy was implemented; the method further includes: Determine a test group and a control group corresponding to the target business strategy. The test group is the user subgroup that was subjected to the target business strategy during the observation period, and the control group is the user subgroup that was not subjected to the target business strategy during the same period. Based on the data of the test group and the control group during the baseline period before the implementation of the target business strategy, a homogeneity test was conducted on the test group and the control group. Provided that the homogeneity test is passed, the actual contribution value of the target business strategy to the business indicators is calculated based on the data of the test group and the control group during the observation period after the implementation of the target business strategy; wherein, the actual contribution value is used to generate the attribution conclusion of the abnormal fluctuation.

10. The method of claim 9, further comprising: In the absence of a control group corresponding to the target business strategy, a similar historical strategy of the target business strategy is matched from the historical strategy library, and the control group corresponding to the similar historical strategy is determined as a substitute control group. Based on the data of the test group during the observation period after the implementation of the target business strategy, and the data of the alternative control group during the historical observation period after the implementation of the similar historical strategy, the estimated contribution value of the target business strategy to the business indicators is calculated. The estimated contribution value is used to generate the attribution conclusion for the abnormal fluctuation.

11. A credit risk attribution device, integrated into a computer system, comprising: The analysis group determination unit is configured to determine the credit user group to be analyzed based on abnormal fluctuations in credit risk indicators in credit business. The dimensionality reduction processing unit is configured to perform dimensionality reduction processing on the original feature set involved in the credit user group to obtain a key feature subset; The labeled data construction unit is configured to construct a labeled dataset based on the feature values ​​of each user in the credit user group on the key feature subset and the corresponding risk labels; the risk labels indicate whether the user is in a preset risk state, and the definition of the preset risk state depends on the credit risk indicators. The path determination unit is configured to determine a decision tree path that conforms to a predetermined standard based on the labeled dataset; The decision tree path includes feature terms and corresponding value conditions, and the predetermined standard includes: the coverage of the decision tree path for users in the predetermined risk state reaches a predetermined threshold. The conclusion generation unit is configured to generate attribution conclusions for the abnormal fluctuations based on the decision tree path.

12. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed in the computer, it causes the computer to perform the method according to any one of claims 1-10.

13. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-10.

Images