Credit approval adjustment method and device based on anti-fact causal atlas

Abstract

The invention discloses a credit examination and approval adjustment method and device based on an anti-fact causal atlas, and relates to the field of financial science and technology, and the method comprises the steps: building a causal atlas based on a current credit examination and approval chain, and obtaining a credit examination and approval adjustment result based on a rule block set, a risk segment set, a strategy version timeline and the causal atlas corresponding to the current credit examination and approval chain; determining a candidate sample set from a case library, generating an anti-fact sample corresponding to each candidate sample based on an abnormal causal path, obtaining an anti-fact sample pair set, splitting a current credit approval chain to obtain a feature group set, a rule block set and an approval conclusion, and obtaining a feature group set, a rule block set and an approval conclusion based on the anti-fact sample pair set; and determining a contribution value of each feature group under each rule block in the rule block set to obtain a layered contribution vector, determining a minimum adjustment entry based on the layered contribution vector, and adjusting the current credit approval chain based on the minimum adjustment entry. The technical problem that the credit approval chain cannot be quickly adjusted in related technologies is solved.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and more specifically, to a credit approval adjustment method and apparatus based on counterfactual causal graphs. Background Technology

[0002] In financial institutions' lending operations, compliance governance solutions typically break down the approval chain into three layers: a separate scoring stage, a rules engine, and manual review, with statistical fairness indicators serving as the core monitoring framework. Current monitoring methods involve segmenting offline batches by model scores and comparing pass rates. If discrepancies exceed preset thresholds, overall corrections are made through model retraining or rule fine-tuning, followed by retesting in the next cycle. This process, presented as a dashboard monitoring system for the risk department, meets the annual compliance report's disclosure requirements for indicators such as the equilibrium rate and rejection ratio.

[0003] To more accurately explain the sources of discrepancies, interpretability algorithms can be introduced at the model layer, such as SHAP (Shapley Additive Explanations, an interpretation method based on Shapley values ​​used to assign contribution values) or LIME (Local Interpretable Model-agnostic Explanations). The system first fixes the trained scorecard or machine learning model, then outputs feature contribution vectors for each sample, and infers potential biases by summing the average weights of visible features. When certain variables related to preset attributes contribute significantly, business personnel may perform feature masking, binning redesign, or coefficient decay to reduce fairness risks.

[0004] In terms of policy-based rule governance, the system primarily relies on white-box visualization and breakpoint backtracking. The rule engine writes thresholds, version numbers, and trigger logs to a configuration repository. When complaints or challenges arise, auditors can query rule snapshots based on timestamps and replay the approval process to replicate the decision-making process. Some systems are equipped with breakpoint sandboxes, allowing for single-point testing of key thresholds without affecting the online environment to observe real-time changes in pass rates and inequality metrics.

[0005] Currently, causal inference is beginning to be applied in the review process. The approach involves establishing structural equations or simplified Bayesian diagrams, using preset attributes as root nodes, and estimating their total and indirect effects on the approval conclusion. Although this can isolate some reasonable differences in business operations, in practice, most implementations still remain at the overall model level and have not yet been refined to the granularity of individual rules or versions.

[0006] The final stage of compliance monitoring typically involves configuring sentinel thresholds. When the system detects deviations in statistics or anomalies in explanatory metrics, it triggers alerts via email, work orders, or dashboards, and records the event number to satisfy audit documentation requirements. The remediation process often involves team meetings to determine responsibilities, followed by submitting model or rule change requests in the next iteration. Once completed, the monitoring loop is closed again.

[0007] In related technologies, current compliance governance often focuses on batch-based statistical monitoring, emphasizing overall indicators such as pass rates and rejection ratios. This often fails to capture instances of different batches with the same risk at critical thresholds or during policy version transitions, and it lacks sufficient awareness of biases in fine-grained rules. While interpretable methods at the model layer can output feature contributions, they default to comparing all causal paths, lacking constraints on the illegal propagation of proxy variables corresponding to preset attributes, resulting in an inability to thoroughly distinguish between reasonable business differences and discriminatory differences. Rule engines rely on manual replay for breakpoint backtracking, with frequent jumps in the retrieval process and the inability to automatically generate control samples, making it difficult to quickly locate specific rule versions. Furthermore, current causal inference practices typically remain at the level of overall model effect estimation, failing to align with regulatory-defined approval process boundaries and policy version lifecycles, thus failing to create an immediately actionable remediation entry point. Sentinel monitoring can only trigger alarms based on preset thresholds, lacking consistent binding with rule-level evidence and responsible parties. While traces are left, source tracing and remediation still heavily rely on manual judgment.

[0008] Therefore, there is an urgent need for an accurate and efficient compliance governance method to enable rapid adjustments to abnormal credit approvals.

[0009] There is currently no effective solution to the above problems. Summary of the Invention

[0010] This invention provides a credit approval adjustment method and apparatus based on counterfactual causal graphs, which at least solves the technical problem of the inability to quickly adjust the credit approval chain in related technologies.

[0011] According to one aspect of the present invention, a method for adjusting credit approval based on a counterfactual causal graph is provided, comprising: constructing a causal graph based on the current credit approval chain, wherein the nodes in the causal graph are abstracted from the process modules on the current credit approval chain, and the directed edges in the causal graph are used to indicate the relationships between nodes; the causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths; and determining a candidate sample set from a case library based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain, wherein the case library includes: multiple approved... Completed credit case samples; based on abnormal causal paths, generate counterfactual samples corresponding to each candidate sample in the candidate sample set, resulting in a counterfactual sample pair set, where each counterfactual sample pair in the counterfactual sample pair set consists of a candidate sample and its corresponding counterfactual sample; decompose the current credit approval chain to obtain a feature group set, a rule block set, and an approval conclusion, and based on the counterfactual sample pair set, determine the contribution value of each feature group in the feature group set under each rule block in the rule block set, resulting in a hierarchical contribution vector; based on the hierarchical contribution vector, determine the minimum adjustment entry point, and adjust the current credit approval chain based on the minimum adjustment entry point.

[0012] Furthermore, the steps for constructing a causal graph based on the current credit approval chain include: abstracting each process module in the current credit approval chain into nodes, wherein each node includes at least: scoring nodes, rule block nodes, manual review nodes, approval conclusion nodes, preset attribute nodes, proxy variable nodes, and business intermediary nodes; determining directed edges based on the business relationships between nodes, and constructing a causal skeleton based on all nodes and all directed edges; determining preset attribute nodes and proxy variable nodes as source nodes, and determining preset paths from the source nodes to the approval conclusion nodes based on the causal skeleton; determining path labels for each preset path based on the business intermediary nodes, and marking the path labels on the causal skeleton to obtain an initial causal graph; mapping the path labels on the initial causal graph to the rule block set and the strategy version timeline, establishing a correspondence between path identifiers, rule identifiers, and strategy version numbers, and writing the correspondence into the initial causal graph to obtain the causal graph.

[0013] Furthermore, based on the rule block set, risk segment set, strategy version timeline, and causal graph corresponding to the current credit approval chain, the steps to determine the candidate sample set from the case library include: mapping the causal graph to the rule block set, risk segment set, and strategy version timeline to determine the breakpoint set for each rule block, wherein each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point; classifying all credit case samples in the case library according to rule blocks and segmenting them according to the strategy version timeline to obtain an index; based on the index and the risk segment set, determining the set of credit case samples located in the same risk segment and with the same credit information, and determining two adjacent credit case samples in the credit case sample set as sample pairs according to time order; filtering all sample pairs based on the approval conclusions of the credit case samples to obtain an initial sample pair set; and processing the initial sample pair set based on the causal graph to obtain an abnormal path descendant feature set and a candidate sample set.

[0014] Furthermore, the step of generating a counterfactual sample corresponding to each candidate sample in the candidate sample set based on the abnormal causal path includes: determining the set of node values ​​of the source node and freezing the reasonable business path on the causal graph; for each candidate sample, replacing the original value of the preset attribute or proxy variable indicated by the source node in the candidate sample with the node value in the set of node values ​​to obtain the replaced candidate sample; deriving the replaced candidate sample on the frozen causal graph to obtain the initial counterfactual sample; for each initial counterfactual sample, verifying the initial counterfactual sample with the breakpoint set, risk segmentation set and strategy version timeline to obtain the counterfactual sample, and updating the abnormal path descendant features of the abnormal causal path affected by the counterfactual sample.

[0015] Furthermore, based on the counterfactual sample pair set, the step of determining the contribution value of each feature group in the feature group set under each rule block in the rule block set, and obtaining the hierarchical contribution vector, includes: establishing a mapping table between feature groups and rule blocks; comparing each counterfactual sample pair in the counterfactual sample pair set with the counterfactual sample as a benchmark to obtain normal features, and freezing all normal features; adding all abnormal path descendant features to the current credit approval chain for testing, and obtaining test results, wherein the test results include at least: the abnormal causal paths affected by each abnormal path descendant feature; constructing a restricted test environment based on the test results; and in the restricted test environment, using a hierarchical algorithm to calculate the contribution value of each feature group under each rule block, and obtaining the hierarchical contribution vector.

[0016] Furthermore, after obtaining the hierarchical contribution vector, the process also includes: aligning the hierarchical contribution vector with the rule block and strategy version timeline to generate an evidence chain; generating a heatmap based on the evidence chain; integrating the rule block and strategy version according to a preset order based on the hierarchical contribution vector to obtain an evidence report; and encapsulating the evidence chain, heatmap, and evidence report to obtain a source tracing report.

[0017] Furthermore, the step of determining the minimum adjustment entry point based on the hierarchical contribution vector includes: determining the rules to be adjusted and the time point of each rule to be adjusted based on the source tracing report; determining the adjustment breakpoint of each rule to be adjusted based on the time point of each rule to be adjusted; and determining the minimum adjustment entry point based on all rules to be adjusted, the time points of all rules to be adjusted, and the adjustment breakpoints of all rules to be adjusted.

[0018] According to another aspect of the present invention, a credit approval adjustment device based on a counterfactual causal graph is also provided, comprising: a construction unit, configured to construct a causal graph based on the current credit approval chain, wherein the nodes in the causal graph are nodes abstracted from each process module on the current credit approval chain, and the directed edges in the causal graph are used to indicate the relationship between nodes; the causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths; and a first determination unit, configured to determine a candidate sample set from a case library based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain, wherein the case library includes: multiple approved credit approval records. The system comprises three parts: a loan case sample generation unit, ...

[0019] Furthermore, the construction unit includes: a first abstraction module, used to abstract each process module on the current credit approval chain into nodes, wherein the nodes include at least: scoring nodes, rule block nodes, manual review nodes, approval conclusion nodes, preset attribute nodes, proxy variable nodes, and business intermediary nodes; a first determination module, used to determine directed edges based on the business relationships between nodes, and to construct a causal skeleton based on all nodes and all directed edges; a second determination module, used to determine the preset attribute nodes and proxy variable nodes as source nodes, and to determine the preset paths starting from the source nodes and pointing to the approval conclusion nodes based on the causal skeleton; a third determination module, used to determine the path labels of each preset path based on the business intermediary nodes, and to mark the path labels on the causal skeleton to obtain an initial causal graph; and a first mapping module, used to map the path labels on the initial causal graph to the rule block set and the strategy version timeline, to establish the correspondence between path identifiers, rule identifiers, and strategy version numbers, and to write the correspondence into the initial causal graph to obtain the causal graph.

[0020] Further, the first determining unit includes: a fourth determining module, used to map the causal graph to the rule block set, risk segment set, and strategy version timeline, and determine the breakpoint set for each rule block, wherein each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point; a first classification module, used to classify all credit case samples in the case library according to the rule blocks and segment them according to the strategy version timeline to obtain an index; a fifth determining module, used to determine the set of credit case samples located in the same risk segment and with the same credit information based on the index and the risk segment set, and to determine two adjacent credit case samples in the credit case sample set as sample pairs according to the time order; a first screening module, used to screen all sample pairs based on the approval conclusion of the credit case samples to obtain an initial sample pair set; and a first processing module, used to process the initial sample pair set based on the causal graph to obtain an abnormal path descendant feature set and a candidate sample set.

[0021] Furthermore, the generation unit includes: a sixth determining module, used to determine the set of node values ​​of the source node and freeze the reasonable business paths on the causal graph; a first replacement module, used to replace the original values ​​of the preset attributes or proxy variables indicated by the source node in each candidate sample with the node values ​​in the set of node values ​​to obtain the replaced candidate sample; a first derivation module, used to deduce the replaced candidate sample on the frozen causal graph to obtain the initial counterfactual sample; and a first verification module, used to verify each initial counterfactual sample with the breakpoint set, the risk segmentation set, and the strategy version timeline to obtain the counterfactual sample and update the abnormal path descendant features of the abnormal causal paths affected by the counterfactual sample.

[0022] Further, the second determining unit includes: a first establishing module for establishing a mapping table between feature groups and rule blocks; a first comparison module for comparing each counterfactual sample pair in the counterfactual sample pair set with counterfactual samples as a benchmark to obtain normal features and freeze all normal features; a first adding module for adding all abnormal path descendant features to the current credit approval chain for testing to obtain test results, wherein the test results include at least: the abnormal causal paths affected by each abnormal path descendant feature; a first construction module for constructing a restricted test environment based on the test results; and a first calculation module for calculating the contribution value of each feature group under each rule block in the restricted test environment using a layered algorithm to obtain a layered contribution vector.

[0023] Furthermore, the credit approval adjustment device also includes: a first alignment module, used to align the hierarchical contribution vector with the rule block and strategy version timeline after obtaining the hierarchical contribution vector to generate an evidence chain; a first generation module, used to generate a heat map based on the evidence chain; a first integration module, used to integrate the rule block and strategy version according to a preset order based on the hierarchical contribution vector to obtain an evidence report; and a first encapsulation module, used to encapsulate the evidence chain, heat map, and evidence report to obtain a source tracing report.

[0024] Furthermore, the third determining unit includes: a seventh determining module, used to determine the rules to be adjusted and the time point of each rule to be adjusted based on the source tracing report; an eighth determining module, used to determine the adjustment breakpoint of each rule to be adjusted based on the time point of each rule to be adjusted; and a ninth determining module, used to determine the minimum adjustment entry point based on all rules to be adjusted, the time points of all rules to be adjusted, and the adjustment breakpoints of all rules to be adjusted.

[0025] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any one of the above-described credit approval adjustment methods based on a counterfactual causal graph.

[0026] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described credit approval adjustment methods based on counterfactual causal graphs.

[0027] In this invention, a causal graph is constructed based on the current credit approval chain. Based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain, a candidate sample set is determined from the case library. Based on abnormal causal paths, counterfactual samples corresponding to each candidate sample in the candidate sample set are generated, resulting in a counterfactual sample pair set. The current credit approval chain is then split to obtain a feature group set, a rule block set, and an approval conclusion. Based on the counterfactual sample pair set, the contribution value of each feature group in the feature group set under each rule block in the rule block set is determined, resulting in a hierarchical contribution vector. Based on the hierarchical contribution vector, the minimum adjustment entry point is determined, and the current credit approval chain is adjusted based on the minimum adjustment entry point, thereby solving the technical problem of the inability to quickly adjust the credit approval chain in related technologies.

[0028] In this invention, the approval process module can first be abstracted into nodes to form a directed causal skeleton, and path attributes can be labeled. Then, by combining rule blocks, risk segments, and strategy version timelines, candidate samples with the same risk but different batches are selected from the case library. Next, based on abnormal causal paths, counterfactual comparisons are generated for each candidate sample to form a compliance difference sample set. Subsequently, the approval chain is decomposed hierarchically according to feature groups, rule blocks, and conclusions. The counterfactual samples are used to calculate the hierarchical SHAP contribution value limited by abnormal paths, and the rule-level impact details are output. Finally, the minimum adjustment entry point of the specific rule threshold is located based on the contribution vector. Through versioned recording and regression verification, accurate and traceable rectification of the credit approval chain is achieved, which effectively improves the automation level and rectification efficiency of algorithm compliance governance, thereby achieving the technical effect of accurately locating and analyzing implicit discrimination. Attached Figure Description

[0029] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0030] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a credit approval adjustment method based on counterfactual causal graphs is shown.

[0031] Figure 2 This is a flowchart of a credit approval adjustment method based on a counterfactual causal graph according to Embodiment 1 of the present invention;

[0032] Figure 3 This is a schematic diagram comparing the prior art with the technology of the present invention according to Embodiment 1;

[0033] Figure 4 This is a schematic diagram illustrating the determination of counterfactual samples according to Embodiment 1 of the present invention;

[0034] Figure 5 This is a schematic diagram illustrating the determination of the hierarchical contribution vector according to Embodiment 1 of the present invention;

[0035] Figure 6 This is a schematic diagram illustrating the determination of the traceability report according to Embodiment 1 of the present invention;

[0036] Figure 7 This is a schematic diagram illustrating the determination of the minimum adjustment entry point according to Embodiment 1 of the present invention;

[0037] Figure 8 This is a schematic diagram of the process for detecting and tracing implicit discrimination in credit approval based on counterfactual causal graphs and hierarchical SHAP value attribution according to Embodiment 1 of the present invention.

[0038] Figure 9 This is a schematic diagram of an optional credit approval adjustment device based on a counterfactual causal graph according to an embodiment of the present invention;

[0039] Figure 10 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0040] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0041] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0042] It should be noted that all related information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and it does not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface. After receiving consent from the aforementioned user or organization, the relevant information is obtained. If the user chooses to refuse, the process proceeds to an expert decision-making process.

[0043] This invention belongs to the field of financial technology and algorithm compliance governance technology, and in particular relates to an automated method that, based on counterfactual causal graphs and hierarchical SHAP attribution, accurately discovers implicit discrimination caused by preset attributes and their proxy variables, locates the responsible rule blocks and generates the minimum rectification path, and forms an automated loop from discovery to tracing, from tracing to rectification, and from rectification to monitoring in the process of bank credit approval chain model scoring to rule engine, rule engine to manual review.

[0044] This invention addresses the challenge of constructing a counterfactual comparison consistent with regulatory policies within a complex credit approval chain encompassing model scoring, rule engines, and manual review. It strictly limits the comparison scope to illegal causal paths of preset attributes and their proxy variables, accurately separating decision differences caused by these paths from reasonable business differences. This allows for precise identification of specific rule blocks and strategy versions, generating auditable and executable evidence and a minimal entry point for rectification. Unlike current solutions that primarily conduct fairness reviews at the level of indicator monitoring, overall interpretability, or white-box replay, this invention first unifies scoring nodes, rule blocks, manual review, preset attributes, and proxy variables into a causal framework. It then marks the propagation direction and level according to the regulatory process boundaries, thus distinguishing reasonable business paths from illegal causal paths in a static graph. This prior routing division naturally converges the subsequent comparison scope to illegal propagation of regulatory concern, avoiding the potential for reasonable business differences to be mixed in with the average contribution distribution across the entire path in traditional SHAP or LIME methods.

[0045] In this invention, regarding sample localization and comparison construction, a rule block-breakpoint-version ternary index is used to simultaneously lock threshold boundaries, risk segments, and policy versions, generating pairs of samples with the same risk within the same neighborhood. This is then combined with graph labels to automatically generate counterfactual comparisons consistent with policies. Compared to breakpoint sandboxes or structural equation replacement of the overall model that rely on manual replay, this method can complete the localization of comparable samples for a single rule threshold in milliseconds, ensuring window consistency and directly meeting regulatory compliance requirements for paired comparability and window traceability.

[0046] In this invention, for the attribution level, a layered SHAP (Restricted Consortium) approach is adopted. This approach integrates a three-layer structure of feature group-rule block-approval conclusion with real-time verification of version visibility and responsibility attribution. Only within the descendant set of illegal paths are candidate features sequentially merged according to gating priority, with subsequent descendants, ensuring that the calculation of marginal increments is strictly limited to the neighborhood of the breakpoint. The resulting contribution vector can be mapped one-to-one to the rule ID (identifier), threshold coordinates, and effective timeline, and directly pushed to the rule engine as the rectification entry point, forming a one-stop closed loop of discovery-source tracing-minimum rectification. Current technologies often only retain balance rate differences at the report level, requiring additional manual analysis to locate specific rules.

[0047] In this invention, by binding the rectification plan to versioned records and performing regression verification on the same neighborhood sample set, closed-loop self-consistency verification and sentinel-style continuous monitoring are achieved, ensuring that every threshold fine-tuning has a traceable, verifiable, and rollback-capable structured trace. This whole-chain design, from graph modeling to limited attribution to minimum rectification, provides compliance departments with an interpretation and execution tool that is completely consistent with the granularity of rules, achieving a fundamental distinction from current batch statistics and overall interpretable methods.

[0048] The present invention will now be described in detail with reference to various embodiments.

[0049] Example 1

[0050] According to an embodiment of this application, an embodiment of a credit approval adjustment method based on a counterfactual causal graph is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0051] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a credit approval adjustment method based on counterfactual causal graphs is shown. Figure 1As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1 The processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions may also be included. In addition, it may include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera, wherein the network interface can be connected to wired and / or wireless networks. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0052] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0053] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the credit approval adjustment method based on counterfactual causal graphs in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned credit approval adjustment method based on counterfactual causal graphs. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0054] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0055] The display can be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0056] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for adjusting credit approval based on counterfactual causal graphs is shown. Figure 2 This is a flowchart of a credit approval adjustment method based on a counterfactual causal graph according to Embodiment 1 of the present invention, as shown below. Figure 2 As shown, the method includes the following steps:

[0057] Step S201: Based on the current credit approval chain, construct a causal graph. The nodes in the causal graph are abstracted from the various process modules on the current credit approval chain. The directed edges in the causal graph are used to indicate the relationship between nodes. The causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths.

[0058] In this embodiment of the invention, on a credit approval chain that includes model scoring, rule engine, and manual review, scoring nodes, rule blocks, manual review nodes, approval conclusion nodes, preset attributes, proxy variables, and business intermediaries are uniformly abstracted into a node set. The direction of directed edges and the propagation range are determined according to regulatory standards to form a causal skeleton. For paths pointing from preset attributes or their proxies to the approval conclusion, a set of reasonable business paths and a set of illegal causal paths are marked. Each path segment is mapped to the corresponding rule block identifier, threshold position, and strategy version, and a causal graph with path labels is output.

[0059] In this embodiment of the invention, the current credit approval chain refers to a series of standardized and automated decision-making processes and systems adopted by financial institutions in the credit approval process. By abstracting the credit approval process, a clear causal relationship model is formed, namely a causal graph. The nodes in this graph represent key components on the approval chain, including scoring nodes, rule blocks, manual review nodes, approval conclusion nodes, as well as preset attributes related to the decision (such as gender), proxy variables (which may indirectly reveal information about protected attributes), and business intermediaries (such as income and debt status). The directed edges in this graph are used to indicate the causal relationships between nodes, such as how scoring affects the decision of the rule engine, or how preset attributes may potentially affect the approval result through proxy variables.

[0060] Here, path labels are used to clearly mark legitimate business paths and abnormal causal paths (i.e., illegal causal paths), ensuring that subsequent analysis focuses on paths that may lead to implicit discrimination. In practice, historical approval data can be analyzed in conjunction with regulatory policies and credit standards to determine which paths reflect legitimate business considerations (such as credit scores) and which paths are abnormal causal paths, i.e., decision-making bases not recognized by regulators, such as approval biases directly or indirectly based on preset attributes.

[0061] Step S202: Based on the rule block set, risk segment set, strategy version timeline and causal graph corresponding to the current credit approval chain, determine the candidate sample set from the case library, wherein the case library includes: multiple credit case samples that have been approved.

[0062] In this embodiment of the invention, the causal graph is mapped to the rule block set, risk segment set, and strategy version timeline to determine the threshold breakpoint and neighborhood buffer range of each rule block. Within each neighborhood, approval records with scores in the same risk segment and consistent products, credit limits, and repayment methods are paired nearest to each other, generating sample pairs of different batches with the same risk. Sample pairs that can be explained by a reasonable business path are then removed, resulting in a candidate sample set. Specifically, based on the rule block set, risk segment set, strategy version timeline, and causal graph, a representative candidate sample set that meets the compliance comparison premise can be selected from a case library containing a large number of approved credit cases. This involves finding sample pairs with different approval results within the same risk segment, and these differences cannot be explained by a reasonable business path. By establishing a ternary index of rule block-breakpoint-version, the comparison is ensured to be conducted under the same risk assessment conditions, improving the accuracy and significance of sample comparison.

[0063] Step S203: Based on the abnormal causal path, generate counterfactual samples corresponding to each candidate sample in the candidate sample set to obtain a counterfactual sample pair set, wherein the counterfactual sample pairs in the counterfactual sample pair set are composed of candidate samples and the counterfactual samples corresponding to the candidate samples.

[0064] In this embodiment of the invention, interventionable nodes and frozen nodes are determined based on the set of illegal causal paths in the causal graph. Preset attributes or their proxies are replaced with compliant alternative values. Descendant features are derived only along illegal causal paths, generating counterfactual sample pairs corresponding to the candidate sample set. Subsequently, breakpoint consistency and neighborhood masks are checked, retaining counterfactual sample pairs that trigger illegal causal paths and have consistent windows. Specifically, by modifying protected attributes and their proxies in the candidate samples, a set of hypothetical approval results is created under the condition that these attribute values ​​change, while other attributes unaffected by abnormal causal paths remain unchanged. This allows for a visual demonstration of how attribute changes affect approval results, especially those changes directly related to implicit discrimination.

[0065] Step S204: The current credit approval chain is split to obtain a set of feature groups, a set of rule blocks, and an approval conclusion. Based on the counterfactual sample pair set, the contribution value of each feature group in the feature group set under each rule block in the rule block set is determined to obtain a hierarchical contribution vector.

[0066] In this embodiment of the invention, for counterfactual sample pairs, the approval chain is split into three layers: feature groups, rule blocks, and approval conclusions. A restricted feature alliance is constructed based on mapping relationships, and candidate features are sequentially incorporated according to gating priority and topological order. Restricted hierarchical SHAP calculation is performed to obtain the marginal contribution vector of each feature group under the corresponding rule block and strategy version, and this vector is written into the rule identifier, threshold position, version number, and approval log reference. Specifically, the current credit approval chain is further subdivided into three layers: a set of feature groups, a set of rule blocks, and approval conclusions. Using the set of counterfactual sample pairs, the contribution value of each feature group under the rule block is calculated, and then summarized to form a hierarchical contribution vector. Here, the contribution value is obtained through restricted hierarchical SHAP calculation, specifically focusing on differences caused by abnormal causal paths, ensuring that it is strictly limited to the range of suspected discrimination, and avoiding reasonable business differences that may be mixed in by traditional global contribution calculations.

[0067] In this embodiment of the invention, the hierarchical contribution vector is used to clarify the degree of responsibility of each feature group in causing approval discrepancies, which helps to accurately locate the root cause of the problem and provide a basis for the formulation of rectification plans.

[0068] Step S205: Based on the hierarchical contribution vector, determine the minimum adjustment entry point, and adjust the current credit approval chain based on the minimum adjustment entry point.

[0069] In this embodiment of the invention, contribution vectors are aggregated in the order of feature groups to rule blocks, and rule blocks to policy versions, generating a rule-level contribution detail and version timeline alignment table. Sample pairings, illegal path labels, hierarchical heatmaps, trigger rule details, and log fragments are then encapsulated as an auditable evidence package. Next, based on the rule-level contribution details, the smallest executable rectification entry point is located. A rectification plan is generated within the neighborhood of the corresponding threshold breakpoint, and regression verification is performed on the same neighborhood sample set. The verified rectification plan is then solidified and linked to the rule identifier, version timeline, and approval logs.

[0070] Specifically, based on the hierarchical contribution vector, the rule blocks and feature groups with the greatest impact and the most need for adjustment can be identified, thus determining the minimum adjustment entry point. By focusing on key threshold adjustments and rule fine-tuning, rather than making comprehensive changes, precise intervention in the credit approval process is achieved. Furthermore, testing the adjustment results in a sandbox environment ensures the safety and effectiveness of the rectification, avoiding the potential risks of blind modifications.

[0071] Figure 3 This is a schematic diagram comparing the prior art with the technology of the present invention according to Embodiment 1, as shown below. Figure 3 As shown, the existing technical process is as follows: First, data sampling and preprocessing: random sampling / resampling, without explicit causal constraints; ordinary interpretation, i.e., global SHAP interpretation: calculating the contribution of all features at once, with different features having equal weight, without stratification; no causal constraints, i.e., making the feature independence assumption: lacking protected attributes (i.e., preset attributes) / illegal path control, easily producing spurious correlation interpretations; manual processing, i.e., manually adjusting parameters one by one: experts fine-tune the thresholds based on experience, which is time-consuming and cannot be repeated; repeated debugging: i.e., the result deviation cannot be traced: compliance / fairness issues are difficult to locate, and rectification is time-consuming. This invention modifies existing technologies to achieve the following innovative process: Causal graph construction: abstracting scores, rule blocks, protected attributes, etc., to form a restricted propagation graph (adding an innovation not found in existing technologies ②: causal graph constraints); Restricted attribution, i.e., restricted hierarchical SHAP: based on the illegal descendant feature alliance, attribution is performed according to "feature group -> rule block -> version" (adding an innovation not found in existing technologies ①: hierarchical + restricted SHAP); Sample comparison, i.e., same-window counterfactual samples: freezing reasonable business paths and generating only comparison samples along illegal paths; Location rectification, i.e., automated minimum rectification: accurately locking threshold breakpoints and impact boundaries, and automatically regressing for verification (adding an innovation not found in existing technologies ③: automated minimum rectification); Closed-loop completion, i.e., closed-loop monitoring and continuous correction: version solidification + sentinel monitoring, real-time early warning and tracking of effectiveness (adding an innovation not found in existing technologies ④: closed-loop monitoring).

[0072] In summary, the approval process module can be abstracted into nodes to form a directed causal skeleton, and path attributes can be labeled. Then, by combining rule blocks, risk segments, and strategy version timelines, candidate samples of different batches with the same risk can be selected from the case library. Next, based on abnormal causal paths, counterfactual comparisons are generated for each candidate sample to form a compliance difference sample set. Subsequently, the approval chain is decomposed hierarchically according to feature groups, rule blocks, and conclusions. The contribution value of hierarchical SHAP limited by abnormal paths is calculated using counterfactual samples, and rule-level impact details are output. Finally, the minimum adjustment entry point of specific rule thresholds is located based on the contribution vector. Through versioned recording and regression verification, accurate and traceable rectification of the credit approval chain can be achieved, effectively improving the automation level and rectification efficiency of algorithm compliance governance, thereby achieving the technical effect of accurately locating and analyzing implicit discrimination.

[0073] To improve the accuracy of constructing the causal graph, in the credit approval adjustment method based on counterfactual causal graph provided in Embodiment 1 of this application, each process module on the current credit approval chain is abstracted as a node. Each node includes at least: a scoring node, a rule block node, a manual review node, an approval conclusion node, a preset attribute node, a proxy variable node, and a business intermediary node. Based on the business relationships between nodes, directed edges are determined, and a causal skeleton is constructed based on all nodes and all directed edges. Preset attribute nodes and proxy variable nodes are determined as source nodes, and based on the causal skeleton, a preset path is determined from the source node to the approval conclusion node. Based on the business intermediary node, a path label for each preset path is determined, and the path label is marked on the causal skeleton to obtain an initial causal graph. The path labels on the initial causal graph are mapped to the rule block set and the strategy version timeline, establishing a correspondence between path identifiers, rule identifiers, and strategy version numbers. This correspondence is then written into the initial causal graph to obtain the causal graph.

[0074] In this embodiment of the invention, the various business process modules in the current credit approval chain—namely, the scoring node (responsible for calculating customer credit scores), the rule block node (encapsulating specific approval logic rules), the manual review node (used for secondary review of automated processing results), the approval conclusion node (outputting the final credit approval decision), the preset attribute node (protected attributes, such as gender), the proxy variable node (potentially indirectly reflecting the characteristics of preset attributes), and the business intermediary node (reflecting the characteristics of customer financial status, credit history, etc.)—are uniformly abstracted into nodes in a graph. These nodes collectively constitute the basic elements of the credit approval chain, with each node representing a key link or influencing factor in the approval process. Based on the node abstraction, directed edges are determined according to the inherent business logic relationships between the nodes. For example, the scoring node influences the decision of the rule block node, and the output of the rule block node influences the result of the approval conclusion node. In this way, a causal skeleton reflecting the causal relationships of the approval process is formed. This skeleton not only depicts the direct connections between nodes but also reveals the indirect influence paths through topological order, laying the foundation for subsequent path analysis.

[0075] In this embodiment of the invention, preset attribute nodes and proxy variable nodes are defined as source nodes in the causal skeleton because these nodes may have a direct or indirect undue influence on the approval conclusion. By traversing the causal skeleton, starting from the source nodes and following all reachable edges to the approval conclusion node, a series of preset paths are determined. These paths may carry implicitly discriminatory information. Specifically, business intermediary nodes are used as a criterion to divide the preset paths into business-reasonable paths and abnormal causal paths. Business-reasonable paths refer to those that pass through the intermediary chain and reflect business considerations such as the customer's true repayment ability; while abnormal causal paths directly or indirectly affect the approval conclusion based on preset attributes or proxy variables and lack business basis. Each path is assigned a path label and marked on the causal skeleton to distinguish the nature of the path, thereby obtaining a preliminary initial causal graph.

[0076] In this embodiment of the invention, to link path labels with the rule implementations in the actual approval process, the preset path labels in the initial causal graph are further mapped to the rule block set and the policy version timeline. This means that each causal path marked as abnormal will be specific to one or more rule blocks, and associated with the effective time range of these rule blocks, i.e., the policy version timeline. Through this mapping, a correspondence is established between path identifiers, rule identifiers, and policy version numbers, ensuring that the graph information can directly guide to the actual rule details and version change records. This correspondence is formally written into the initial causal graph, forming a causal graph that includes regulatory policies, business logic, and rule implementation details.

[0077] For example, constructing the causal graph S1 includes the following steps:

[0078] S11. The scoring node M, rule block set K, manual review node J, approval conclusion node Y, preset attribute S, proxy variable set Q, and business intermediary set U are uniformly abstracted into a node set V. A directed edge set E is defined based on the influence relationships with business interpretation, forming a causal skeleton G (containing V and E). The direction and propagation range of the edges are determined using regulatory standards and internal process boundaries. Edge types are hierarchically labeled as scoring propagation, rule triggering, review and adjudication, and business intermediary, clarifying the triggering location and propagation restrictions for each type of edge; potential connections lacking business basis are not established. S and Q are labeled as sources, and M, K, J, and Y are labeled as affected nodes, setting a propagation boundary D. b Used to limit the spread of subsequent interventions, it outputs an initial skeleton and node dictionary for path type determination.

[0079] S12. On G constructed in S11, each directed path from S or Q to Y is determined according to topological order. Using compliance criteria and credit policies, the intermediate chains in U reflecting solvency, debt levels, repayment history, income stability, and asset-liability structure are marked as the set of reasonable business paths L. + ; Paths that directly affect M, K, J, and Y without the aforementioned intermediary chain are marked as the set of illegal causal paths L. - For each path, a type label, allowed propagation boundary, and trigger condition reference are written. Priority is given to recording the shortest propagation chain and extension chain related to threshold breakpoints and version switching, forming a path label table and boundary constraint list, which serve as the sole criterion for subsequent counterfactual and attribution analyses.

[0080] S13. Project the path type and boundary constraints generated in S12 onto the rule implementation, and L - With L + The path segment, rule ID, version number, and log reference are mapped to the threshold position, policy version timeline, and approval log field of rule block K, respectively, establishing a correspondence. Auditable identifiers and responsible entities are bound to nodes and edges, using unified naming and version sequence management for fixed mapping. Breakpoint locations, version activation windows, and rule trigger records are written into the graph as evidence anchors, outputting a causal graph G with path type labels. * And the mapping table, which serves as the sole basis for generating policy-consistent counterfactuals and limiting the scope of attribution participation.

[0081] In this embodiment, a precise connection is achieved from abstract business logic to specific rule implementations, ensuring that subsequent implicit discrimination detection based on counterfactual analysis and hierarchical SHAP attribution can be accurate down to the individual rule block and version level. This detailed graph construction approach, extending to the operational level, makes compliance reviews of the credit approval chain more precise, enabling timely identification and location of potentially discriminatory rules. This effectively guides rectification actions with minimal scope, improving the fairness and transparency of the financial approval process. Simultaneously, by managing the correspondence between versioned path tags and rules, abnormal trends in strategy evolution can be tracked, providing a solid foundation for continuous rectification and monitoring, ensuring that the credit approval algorithm always adheres to the principle of fairness and meets regulatory requirements.

[0082] To improve the accuracy of determining the candidate sample set, in the credit approval adjustment method based on counterfactual causal graphs provided in Embodiment 1 of this application, the causal graph is mapped to a set of rule blocks, a set of risk segments, and a strategy version timeline to determine the breakpoint set for each rule block. Each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point. All credit case samples in the case library are categorized according to rule blocks and segmented according to the strategy version timeline to obtain an index. Based on the index and the risk segment set, a set of credit case samples located in the same risk segment and with the same credit information is determined, and two adjacent credit case samples in the credit case sample set are determined as sample pairs according to the time order. Based on the approval conclusion of the credit case samples, all sample pairs are screened to obtain an initial sample pair set. Based on the causal graph, the initial sample pair set is processed to obtain an abnormal path descendant feature set and a candidate sample set.

[0083] In this embodiment of the invention, firstly, the causal graph is mapped to a set of rule blocks, a set of risk segments, and a timeline of strategy versions to establish a bridge between the theoretical model and actual operation, ensuring that subsequent operations can accurately pinpoint specific rules, risk levels, and version change points. By marking the ID of the rule block, its effective version number, and timeline on the graph, every rule change in the approval process and its potential impact can be tracked. Then, a set of breakpoints is determined, covering the thresholds, segment boundaries, and version switching points in each rule block. Thresholds refer to key numerical points in the rule block used for decision-making; for example, a credit score below a certain value will automatically result in rejection. Segment boundaries are defined based on the risk segment set to distinguish different risk levels, with each segment potentially corresponding to a different approval strategy. Version switching points record the precise time of rule or strategy changes, which is crucial for tracking the evolution of approval standards. By identifying these breakpoints, the key moments of decision changes can be precisely located in subsequent analysis.

[0084] In this embodiment of the invention, credit case samples in the case library are classified and segmented according to the aforementioned mapping, forming an index ordered by rule blocks and strategy version timelines. This index not only facilitates quick access to samples but also enables refined management of cases according to chronological order and risk level. Based on this index and risk segment set, credit cases located in the same risk segment, i.e., those with similar credit risk assessment results, are selected. It is ensured that these cases maintain consistency in credit information such as product type, loan amount, and repayment method to guarantee the fairness and comparability of sample comparison. In the selected credit case sample set, adjacent samples are identified as sample pairs based on chronological order. The consideration of chronological order ensures that sample pairs can reflect approval differences under the same approval standards; such sample pairs form the basis for subsequent analysis of implicit discrimination issues.

[0085] In this embodiment of the invention, the generated initial sample pair set is preliminarily screened. By comparing the approval conclusions of each sample pair, sample pairs with differing approval results can be identified, which may indicate implicit discrimination in the approval process. Subsequently, based on causal graphs, each sample pair is analyzed in depth to identify the set of descendant features that influence the approval conclusions through mediating features along anomalous causal paths (i.e., paths not approved by regulators). These features may directly or indirectly reflect information about protected attributes. Finally, a candidate sample set is determined. This set contains credit case sample pairs whose differences in approval results cannot be explained by reasonable business paths and are suspected of implicit discrimination. This set serves as direct input for subsequent counterfactual sample generation and hierarchical SHAP attribution analysis, ensuring that subsequent analysis focuses on potential discriminatory approval differences rather than normal business changes or random fluctuations.

[0086] For example, determining the candidate sample set S2 includes the following steps:

[0087] S21. The causal graph G output by S1 * Mapping the rule block set K, the scoring card risk segment set R, and the strategy version timeline T, we determine the breakpoint set B (including thresholds, segment boundaries, and version switching points) for each rule block. For each breakpoint in B, a left and right buffer configuration W is used to describe the neighborhood range and effective interval. Approval records are merged according to rule block K and segmented according to version interval T, creating a rule block-breakpoint-version ternary index I. Visible fields and call entry points are marked on index I, fixed as unique sampling boundaries and access points for paired positioning of the same risk, for direct reference in subsequent steps.

[0088] S22. Filter the records within each rule block-breakpoint-version window of Index I according to risk segment R, locking in a set of records with the same score segment, consistent credit type, consistent credit limit, and consistent repayment method. Form an ordered sequence of records within the set according to the applicant's subject and time order, and generate a sample pair set P using a proximity pairing strategy. For each pair of records, read the approval conclusion Y and mark it as either approved-rejected or rejected-approved, simultaneously writing the triggered rule block and breakpoint references, and outputting a list of initial screening sample pairs of the same risk but different batches as input for illegal path verification.

[0089] S23. Input the sample pair set P into G one pair at a time. * The path label table allows for the identification of differences based on the set of reasonable business paths, L. + The explained sample pairs are stripped; for those that still exist along the illegal causal path set L - The transmitted sample pairs are preserved. For the preserved sample pairs, the triggering rule block K and the associated illegal path descendant feature set F are extracted. illegal The reference relationships between sample pairs, rule blocks, versions, path segments, and feature sets are solidified. These results are registered as a candidate set C, serving as the sole input caliber and boundary for generating S3 policy consistency counterfactual data.

[0090] In this embodiment, a direct link from the theoretical model to actual samples is effectively constructed. Through precise classification, sorting, and filtering, credit cases most likely reflecting implicit discrimination can be quickly identified. This mapping and indexing based on rule blocks and policy version timelines not only improves analytical efficiency but also ensures the accuracy and relevance of the analysis. Finally, by generating a candidate sample set and anomaly path descendant feature set, a clear direction is provided for subsequent steps, enabling the identification and correction of implicit discrimination to be conducted more scientifically and systematically, thereby greatly improving the fairness and compliance of the credit approval process.

[0091] To improve the accuracy of generating counterfactual samples corresponding to each candidate sample in the candidate sample set, the credit approval adjustment method based on counterfactual causal graph provided in Embodiment 1 of this application determines the set of node values ​​of the source node and freezes the reasonable business paths on the causal graph. For each candidate sample, the original values ​​of the preset attributes or proxy variables indicated by the source node in the candidate sample are replaced with the node values ​​in the set of node values ​​to obtain the replaced candidate sample. The replaced candidate sample is then used to deduce the initial counterfactual sample in the frozen causal graph. For each initial counterfactual sample, the initial counterfactual sample is checked against the breakpoint set, the risk segmentation set, and the strategy version timeline to obtain the counterfactual sample, and the abnormal path descendant features of the abnormal causal paths affected by the counterfactual sample are updated.

[0092] In this embodiment of the invention, for each source node, a set of node values ​​is constructed, containing all possible values ​​of the source node. It is important to note that for preset attributes, these values ​​may come from a list of compliant alternative values, ensuring that the replacement process does not introduce new biases or illegal paths. Before conducting counterfactual analysis, the reasonable business paths on the graph are frozen. This means that any feature derivation or approval process calculation along these paths will remain unchanged and will not be affected by subsequent intervention. This ensures that the analysis process focuses only on changes in abnormal causal paths, avoiding the confusion of results by reasonable business differences. Then, for each sample in the candidate sample set, the original value of the preset attribute or proxy variable indicated by the source node in that sample is replaced with a value from the set of node values. This replacement process is precise down to each abnormal causal path; that is, each replacement attempts to simulate a hypothetical scenario: what impact would a change in the value of the protected attribute or its proxy variable have on the credit approval process? The replaced sample, i.e., the replaced candidate sample, is derived in the causal graph along the same path as the original sample. However, the reasonable business paths in the graph are frozen at this point, and calculations are only performed along illegal paths to generate the initial counterfactual sample. The initial counterfactual sample is then checked against the breakpoint set, risk segment set, and strategy version timeline to ensure that the approval background remains consistent with the original candidate sample and still conforms to the current rules and strategy versions of the credit approval chain. This verification process includes checks on the validity of the sample's risk segment, version, and whether it is in the neighborhood of a breakpoint. If the initial counterfactual sample passes these checks, it is officially confirmed as a counterfactual sample. Furthermore, the comparison between each pair of counterfactual samples and the original sample directly reveals the influence boundary of the abnormal path descendant features. If the approval result of the counterfactual sample differs from the original sample, and this difference can only be explained by illegal paths, then the abnormal path descendant feature set of the abnormal causal path related to that counterfactual sample is updated. This update to the feature set means that more features have been identified that are directly or indirectly affected by protected attributes. These features may carry implicit discrimination information and need to be given special attention in subsequent governance steps.

[0093] For example, determining the counterfactual sample S3 includes the following steps:

[0094] S31. Set the set of compliant alternative values ​​A of the preset attribute S and its proxy set Q. S With causal graph G * Alignment, determining the set L of illegal causal paths - The propagation boundary. For the set of reasonable business paths L... + The intermediaries and outcomes are marked as frozen, preventing alteration during subsequent interventions. The set of propagable nodes V is then... prop The set of nodes prohibited from propagation, V forbAllowed intermediary set U allow Number and describe the interventions to form intervention configuration table C. int This serves as the sole source of information and reference for subsequent counterfactual generation, and it fixes the mapping relationship between rule blocks and versions.

[0095] S32, Configure the intervention of S31 as C int The candidate set C in S2 is applied, and for each original sample x, a compliant alternative value set A is used. S Replace S with the specified value in L. Only along L - In V prop Internally, the relevant descendants are derived step by step, maintaining L + The mediators and outcomes remain unchanged by the frozen marker. For the generated counterfactual sample x... cf Pairing the sample with the corresponding original sample x, the anchoring information of the sample pair-rule block-version-path segment is registered, and a set of comparable counterfactual sample pairs P is output. cf .

[0096] S33, P cf Consistency checks are performed with the breakpoint set B, risk segment set R, and version timeline T, limiting the comparison window to the breakpoint neighborhood buffer configuration W. Sample pairs exceeding the neighborhood or not conforming to regulatory standards are removed; those still acceptable are handled by L. + Excluding sample pairs that explain the discrepancies. Only sample pairs that meet the criteria for illegal path triggering and have consistent windows are retained, forming a list of compliant counterfactual pairs. It serves as a direct input to the attribution step and is kept in correspondence with log references.

[0097] S34, G * The illegal path labels in the image are projected onto the feature layer, and the changes caused by S are processed through L. - The descendant features of the influence are extracted as the unique participation range F illegal For those belonging to L + The associated feature sets the frozen set F freeze Subsequent attribution is prohibited. A fixed mapping is established between sample pairs, rule blocks, versions, path segments, and features, and the output F is determined. illegal With F freeze Its label table serves as a limiting input for the hierarchical attribution step, ensuring that subsequent comparisons proceed only along illegal causal paths.

[0098] Figure 4 This is a schematic diagram illustrating the determination of counterfactual samples according to Embodiment 1 of the present invention, as shown below. Figure 4As shown, the inputs are: a candidate sample set Cpair and an illegal causal graph Γillicit; defining the intervention scope, i.e., determining the intervention configuration: establishing a compliant alternative list for protected attributes and proxy variables, freezing reasonable business path nodes, and clarifying the scope of propagation and prohibition; applying alternative values, i.e., generating counterfactual samples: replacing the alternative values ​​of protected attributes / proxy in each record, deriving descendant features only along the illegal causal path, outputting counterfactual samples Cf and pairing them with the original samples Co; producing Cf / Co, and checking the consistency between the window and the version: verifying the visibility of the breakpoint neighborhood window and the policy version, and removing sample pairs that exceed the window or can be explained by the reasonable business path; after checking the sample pairs, performing feature mapping and generating a frozen set: extracting the illegal descendant feature Ωillicit, labeling the frozen set Ωfreeze, and establishing a fixed mapping of "sample pair -> feature -> path segment".

[0099] In this embodiment, counterfactual samples directly reflecting the impact of abnormal causal paths are generated through precise node value replacement and rigorous process derivation. Simultaneously, verification with actual approval process rules ensures the compliance and validity of these samples, avoiding misleading analysis due to changes in background conditions. The updated set of abnormal path descendant features more comprehensively covers potential discriminatory effects, providing a more solid foundation for subsequent hierarchical SHAP attribution analysis and ensuring the accuracy and relevance of the analysis results. Thus, not only is the accuracy of implicit discrimination identification in the credit approval chain improved, but direct evidence and guidance are also provided for subsequent rectification work, achieving automated and refined operations from discovery to rectification, significantly improving the efficiency and transparency of algorithmic compliance governance in financial institutions.

[0100] To improve the accuracy of determining the hierarchical contribution vector, in the credit approval adjustment method based on counterfactual causal graphs provided in Embodiment 1 of this application, a mapping table between feature groups and rule blocks is established; using counterfactual samples as a benchmark, each counterfactual sample pair in the counterfactual sample pair set is compared to obtain normal features, and all normal features are frozen; all abnormal path descendant features are added to the current credit approval chain for testing to obtain test results, wherein the test results include at least: the abnormal causal paths affected by each abnormal path descendant feature; based on the test results, a restricted test environment is constructed; in the restricted test environment, a hierarchical algorithm is used to calculate the contribution value of each feature group under each rule block to obtain the hierarchical contribution vector.

[0101] In this embodiment of the invention, firstly, a mapping relationship is established between feature groups (feature sets merged according to business logic or semantics) and rule blocks (specific logical units in the approval process) in the credit approval chain, forming a detailed mapping table. The feature groups are divided according to their business functions, such as repayment ability and debt structure, while the rule blocks represent specific logical judgments in the approval process. The mapping table ensures that changes in each feature group can be directly mapped to the decision-making process of the corresponding rule block, facilitating subsequent contribution analysis and rectification positioning. Then, using counterfactual samples as a benchmark, each pair of samples in the counterfactual sample set is carefully compared. Through this process, normal features that remain unchanged even when preset attributes or proxy variables change can be identified. These features are usually associated with business intermediary nodes. All features identified as normal are frozen, that is, they are treated as constants in subsequent analysis and simulation, and do not participate in mutation or calculation, thereby ensuring that the focus of the analysis is on abnormal causal paths and avoiding interference from reasonable business paths. Afterwards, all identified abnormal path descendant features are added to the current credit approval chain for testing. The purpose of this test is to explore how each feature influences anomalous causal paths and how changes to these features alter the approval outcome. The test results should at least include the specific anomalous causal paths influenced by each anomalous path's descendant features, providing precise clues for further problem localization. Based on the preceding tests and analysis, a restricted test environment was constructed, allowing only changes and derivations of features along illegal paths, ensuring the purity and directness of the analysis. In this restricted environment, a layered algorithm (such as the Restricted Layered SHAP algorithm) was used to calculate the contribution value of each feature group under each rule block, outputting a layered contribution vector. This vector clearly demonstrates the degree of influence of each feature group on the approval conclusion in different rule blocks, especially those feature groups related to anomalous causal paths.

[0102] For example, determining the hierarchical contribution vector S4 includes the following steps:

[0103] S41. The approval chain is split into three layers based on business endpoints: feature group, rule block, and final approval. Model features are semantically grouped according to repayment ability, debt structure, and account behavior to form a feature group list, and a one-to-one mapping is established between these feature groups and the triggered rule blocks K. The visibility of the mapping is checked using the strategy version timeline T and the threshold breakpoint set B, and items not in the current version are removed. A hierarchical configuration and mapping table Π is generated using a unified numbering system. The hierarchical scope serves as the sole reference for restricted attribution and aggregation.

[0104] S42. Match each original sample x with its counterfactual sample x. cf Pairing is performed within the same rule block - breakpoint - version window, with x cf Set as the unique baseline; for the causal graph G *Marked as the set of reasonable business paths L + The variables and results are frozen and do not participate in the interpretation of differences; the comparison window is limited to the buffer range W of the rule block K triggered by the sample and its threshold neighborhood B; the list of sample pairs whose baselines have been replaced and are ready to be frozen is output and passed to the restricted coalition construction along with the mapping table Π.

[0105] S43, L - The sample pairs projected onto the S42 viewing window yield the illegal descendant feature set F. illegal With frozen set F freeze Combine the mapping table Π, the threshold breakpoint set B, and the version timeline T to calculate the visibility mask and responsibility attribution for the current rule block K; for F illegal Candidate sequences {f1, ..., f2} are generated in a topological order prioritizing gating and following descendant sequences. m}. Adopting from x cf The unidirectional merging method towards x evaluates candidates sequentially, merging only those that are visible and whose prerequisites are met. For candidates visible across blocks, responsibility-first attribution is applied, incorporating only the current K. Each merging generates a restricted federation, recording the source path segment and version reference. All federations are output as a list C in the generation order, and the mapping between federations, rule blocks, and versions is fixed, serving as the sole input for the restricted hierarchical SHAP.

[0106]

[0107] Among them, C t For the restricted alliance after the t-th merger; f t F represents the t-th candidate feature generated according to gating priority and topological order; illega1 For the feature set of descendants of illegal paths; F freeze For the frozen feature set; ρ(f t K) is the responsibility attribution predicate, and the feature assigned to the current rule block takes a value of 1; v(f t T) is the version visibility predicate, with the current version visibility value being 1; π(f) t C t ) is a predicate that satisfies the condition, indicating that all candidate gates and their parent nodes have already been satisfied in C. t The value is 1; K is the current rule block; T is the current policy version.

[0108] S44, let (x, x cf Alliance List With the mapping table Π as the only input; x cf Set as the unique baseline; within the threshold neighborhood W of the current rule block K, perform mask switching sequentially in order of C, adding only candidate features ft, ensuring comparison along L. - Expand. For each step, first check the version visibility v(f)t ,T) and window mask χ B If (·; K, W) is not satisfied, skip it; if satisfied, update C in the order of gating priority, followed by descendants. t The marginal increment is obtained by calling the approval chain output function Ψ(·) on both the pre-update and post-update states; the increment is assigned to K according to the feature group and rule block path of Π, and responsibility priority is applied to cross-block visible features; the joint impact within the same group is conservatively distributed according to the generation order, without crossing the frozen set F. freeze The summation of the increments of each feature group within K yields H. k For all triggered K, calculate and collect the contribution vector H in sequence, while retaining the reference chain from alliance to group, group to rule block, and rule block to version as the direct input of S45.

[0109]

[0110] Among them, H k Contribution to rule block K; G k G is the set of feature groups mapped from Π to K; g Let m be the feature set of feature group g; m be the alliance length; f be the feature set of feature group g. t C is the t-th merged feature; t For the t-th step alliance; z cf Baseline sample vector; Indicated by C t Coverage z cf Assignment operation; ρ(f t K) is the predicate of responsibility attribution; v(f) t , T) is the version-visible predicate; χ B (·; K, W) is the breakpoint neighborhood window mask; Ψ(·) is the approval chain output function; Π is... H is the contribution vector composed of each K.

[0111] S45. Align H with rule ID, threshold position, policy version, and effective time. Write the responsible entity and approval log reference number for each contribution. Generate a contribution details table in the order of feature group to rule block, rule block to version, and version to time, and fix the binding relationship between each increment and its triggering background. Execute masking for entries that are not visible in the current version and mark the reason. Output contribution details with rule and version tags as evidence solidification and the basis for the chain of responsibility.

[0112] S46. Generate a heatmap and text description of the stratified contributions according to the hierarchical scope of Π, highlighting L. -Related contributions; sample pairing, trigger rules, threshold positioning, version timeline and approval log references are imported into the same evidence package and encapsulated with unified naming and version number; traceable fields and reference verification items are written into the evidence package to ensure that it can be directly received by internal audit and compliance; auditable hierarchical contribution vectors, heat maps and evidence package indexes are output for minimum scope rectification calls.

[0113] Figure 5 This is a schematic diagram illustrating the determination of the hierarchical contribution vector according to Embodiment 1 of the present invention, as shown below. Figure 5 As shown, the input is: counterfactual sample pairs Cf / Co and mapping table M, fixed. The hierarchical information is prepared, and a hierarchical structure and mapping table are constructed: model features are semantically grouped into feature groups, and mapped one-to-one with rule blocks, version timelines, and threshold breakpoints, outputting the restricted hierarchical configuration L; if the configuration is complete, baseline freezing and window limitation are performed: Co is set as the unique baseline within the determined neighborhood, business reasonable path variables are frozen, and the current rule block and version are limited within the comparison window; after freezing the window, restricted alliance generation is performed: candidates are sorted based on the illegal descendant feature set Omegaillicit, and merged in order of "gating priority, descendant later" to form an alliance sequence Lambda and bind responsibility attribution; if mapping reference is used, restricted alliance generation is performed directly; alliance Lambda: restricted hierarchical SHAP calculation and attribution are performed: each time it is merged, the approval chain output function is called to calculate the marginal increment and map it to "feature group -> rule block -> version", summarize the contribution vector Phi and retain the reference chain.

[0114] In this embodiment, features and rules in the credit approval chain are precisely mapped, effectively separating the impact of reasonable business paths from abnormal causal paths. Freezing normal features reduces unnecessary computational complexity, while testing the features of descendants of abnormal paths directly addresses potential discrimination issues. The resulting hierarchical contribution vector not only provides a clear direction for rectification but also enhances the transparency and fairness of the credit approval process, ensuring that every approval decision can withstand rigorous compliance review.

[0115] To improve the accuracy of determining the source tracing report, in the credit approval adjustment method based on counterfactual causal graph provided in Embodiment 1 of this application, after obtaining the hierarchical contribution vector, the hierarchical contribution vector is aligned with the rule block and strategy version timeline to generate an evidence chain; based on the evidence chain, a heat map is generated; based on the hierarchical contribution vector, the rule block and strategy version are integrated according to a preset order to obtain an evidence report; the evidence chain, heat map, and evidence report are encapsulated to obtain the source tracing report.

[0116] In this embodiment of the invention, firstly, the data in the hierarchical contribution vector is precisely aligned with the specific implementation and policy version timeline of the rule block. This means that the contribution value of each feature group recorded in the hierarchical contribution vector under a specific rule block must be closely linked to the performance and changes of that rule block in different policy versions, forming a one-to-one correspondence. This alignment process ensures an accurate understanding of the role of each feature in a specific rule and version, as well as the trend of its contribution to the approval result, providing a solid data foundation for subsequent evidence generation and report preparation. Through the above alignment, a clear chain of evidence can be constructed. This chain of evidence connects the contribution information and change trajectory between each feature group, rule block, and policy version, including but not limited to how the feature group affects the decision of the rule block, and in which versions the rule block was adjusted, leading to the mitigation or aggravation of discriminatory consequences. The generation of the chain of evidence not only provides structured data support for subsequent analysis but also simplifies the audit process to a certain extent because it can transform complex attribution analysis results into easily understandable chains of responsibility relationships, facilitating tracking and review. Then, based on the generated chain of evidence, a heatmap is created using visualization techniques to intuitively display the changes in contribution between feature groups, rule blocks, and policy versions. The color intensity of the heatmap represents the magnitude of the contribution value. Different color levels help quickly identify which parts require special attention—those areas with high contribution values ​​or dramatic changes. The heatmap allows for the rapid location of potential discriminatory rules or versions, providing direct visual clues for subsequent rectification actions. Then, based on the data in the hierarchical contribution vectors, the rule blocks and their associated policy versions are organized and integrated in a predetermined order to form a detailed evidence report. This report not only includes the specific values ​​of the contribution vectors but also combines information from the evidence chain and the heatmap, providing a comprehensive analysis of the origin, scope, and trends of discriminatory issues. To enhance the report's practicality and auditability, all key evidence, including the evidence chain, heatmap, and detailed textual analysis, is encapsulated, ultimately outputting a source tracing report. This report not only summarizes the identified problems but also provides tools and guidelines for tracing the root causes and monitoring the rectification process, significantly improving the efficiency of internal compliance reviews and external supervision within financial institutions.

[0117] For example, determining the traceability report S5 includes the following steps:

[0118] S51. Aggregate the hierarchical contribution vector H output from S4 according to the mapping table Π, performing feature grouping to rule block, rule block to policy version in that order; assign each increment to the rule IDK, version number T, and responsible subject fields, write the position, direction, and trigger time of the threshold breakpoint B, and associate it with the sample pair (x, x). cf() and approval log reference number; for duplicate endpoints across blocks or versions, the attribution criteria of current visibility priority and first trigger priority are used to complete deduplication;

[0119] Generate a rule-level contribution details and version timeline alignment table, fixing the correspondence between contribution → rule → version → log, which serves as the sole input and access point for evidence encapsulation.

[0120] S52. Package the details of S51 into a counterfactual evidence package E according to the audit scope. pkg : Include sample pairings (x, x) cf ), illegal path tags, hierarchical heatmaps, trigger rule details, version timelines, and approval log snippets referenced; generate threshold location pages for each rule ID, marking the coordinates of B and the version switch point T; establish cross-indexes for sample pairs, rules, versions, and logs, and unify naming and version sequence management; output a source tracing report R, and combine E pkg The index of R is registered in the platform directory for direct access by minimal rectification and review processes.

[0121] Figure 6 This is a schematic diagram of determining the traceability report according to Embodiment 1 of the present invention, as shown below. Figure 6 As shown, the input includes a hierarchical contribution vector and mapping table, containing rule ID, threshold coordinates, policy version, and log references. If the input is aggregated, rule-level alignment and aggregation are performed: contributions are summarized according to "feature group -> rule block -> version", the threshold breakpoint position and trigger time are written, and duplicates are generated to produce a rule-level contribution detail table. If it is a detail table, evidence package encapsulation is performed: rule details, sample pairings, heatmaps, path labels, and log fragments are encapsulated together, and a "sample pair -> rule -> version -> log" cross-index is established, outputting a source tracing report and evidence package index. If it is a metadata reference, evidence package encapsulation is performed directly. The evidence package outputs: rule-level contribution detail table + evidence package Epkg + source tracing report Rsrc, which can be directly called by minimal rectification and auditing. If it is rule detail verification, it is output directly.

[0122] This embodiment achieves in-depth analysis and precise identification of discriminatory issues in the credit approval chain. The construction of the evidence chain clarifies responsibility, the heatmap visually presents the severity of the problem, and the final source tracing report integrates all analysis results and evidence into a complete document, facilitating financial institutions' rectification planning and implementation. This not only improves the accuracy of compliance governance but also promotes transparency and efficiency, providing crucial technical support for achieving fairness and impartiality in the credit approval process. Through precise evidence collection and report preparation, financial institutions can better understand and address algorithmic bias issues, ensuring that credit services are fair to all applicants while also meeting regulatory requirements for transparency in algorithmic decision-making.

[0123] To improve the accuracy of determining the minimum adjustment entry point, in the credit approval adjustment method based on counterfactual causal graph provided in Embodiment 1 of this application, the rules to be adjusted and the time point of each rule to be adjusted are determined based on the source tracing report; the adjustment breakpoint of each rule to be adjusted is determined based on the time point of each rule to be adjusted; and the minimum adjustment entry point is determined based on all rules to be adjusted, the time points of all rules to be adjusted, and the adjustment breakpoints of all rules to be adjusted.

[0124] In this embodiment of the invention, firstly, based on the analysis of the evidence chain, heatmap, and hierarchical contribution vector provided in the source tracing report, rule blocks that unfairly impact the approval conclusion at specific points in time can be identified and marked as rules to be adjusted. The determination of rules to be adjusted is based on the prominent abnormal path descendant features and rule block contribution values ​​in the evidence report, accurately locating the specific approval logic unit causing the problem. Simultaneously, the time point of each rule to be adjusted is recorded, i.e., its effective date on the strategy version timeline. This helps to understand the dynamic changes in approval standards and provides a time dimension basis for subsequent breakpoint analysis. Then, for each identified rule to be adjusted, breakpoints to be adjusted are identified based on its time point and the strategy version timeline. Breakpoints typically refer to thresholds or boundary points where approval standards or rules undergo significant changes, including events such as threshold changes, rule structure updates, or strategy version switches. By analyzing the differences in the contributions of rules across different versions, especially those closely related to the characteristics of anomalous path descendants, we can accurately identify which breakpoints led to unfair approval results. This not only identifies problematic rules but also pinpoints the specific rule changes that triggered the problem, providing precise spatial coordinates for defining the minimum adjustment entry points. Subsequently, a comprehensive analysis of all rules to be adjusted, the time points of all rules, and the breakpoints to be adjusted for all rules allows us to determine a set of minimum adjustment entry points. These entry points are considered the most efficient starting points for adjusting the approval process to eliminate or mitigate the impact of implicit discrimination. This may involve fine-tuning thresholds, minor modifications to the rule structure, or partial updates to the policy version, but the goal is always to correct the problem with minimal impact.

[0125] In this embodiment of the invention, the determination of the minimum adjustment entry point can be based on the following principles: ensuring that the rectification measures directly target the impact of the characteristics of the descendants of abnormal paths; minimizing disturbances to reasonable business paths; and considering the cost and feasibility of rule adjustments. By focusing on these minimum adjustment entry points, financial institutions can efficiently implement rectification measures, avoiding the waste of resources and business interruptions caused by large-scale restructuring of approval processes.

[0126] For example, determining the minimum adjustment entry S6 includes the following steps:

[0127] S61. Align the rule-level contribution details of S5 with the version timeline T item by item, and locate the smallest executable rectification entry point according to the reference chain of feature groups, rule blocks, and versions; for each entry point, locate the threshold of the corresponding rule block K to the specific location of the breakpoint set B, and limit the rectification to be implemented only in the neighborhood of the breakpoint; for the illegal descendant feature set F involved in the trigger source. illegal A thorough review was conducted, prioritizing threshold fine-tuning and scope narrowing, followed by blocking or downgrading proxy features; business-reasonable path variables were kept frozen and excluded from the action scope; a rectification plan record was generated, including rule ID, threshold coordinates, version effective window, responsible entity and approval number, as well as action sequence and upper limit of magnitude, impact boundary and execution window, as the input for regression verification.

[0128] S62. Apply the rectification plan from S61 to the set of sample pairs in the breakpoint neighborhood formed in S2, and perform verification according to the planned order within the corresponding version window; for the set of reasonable business paths L... + The relevant variables remain frozen, only for (x, x) cf The approval conclusion is compared with that of the previous one to determine whether the changes in the trigger rule block K and the threshold position B have eliminated the problem of different batches with the same risk. For cases where the problem has not been eliminated or new deviations have been generated, the reasons are recorded and the threshold fine-tuning, scope narrowing and proxy feature control are revised to the minimum extent according to the established priority. The rectification items that have passed the verification are solidified into versioned records, and reference relationships are established with rule ID, version timeline T and approval log. The rectification plan and verification result P are output for direct use by the sentinel monitoring in the future.

[0129] Figure 7 This is a schematic diagram of determining the minimum adjustment entry point according to Embodiment 1 of the present invention, as shown below. Figure 7 As shown, the inputs are: rule-level contribution details, version timeline, and Sentinel monitoring interface metadata; the entry point is located, and a minimum executable rectification plan is generated: precisely locate the rule block and threshold breakpoint neighborhood, design the action sequence, amplitude upper limit, and impact boundary, and output a structured rectification plan; the rectification plan undergoes regression verification and iterative adjustment: the rectification plan is executed within the sandbox window, and if the differences are not eliminated, it is fine-tuned with the minimum amplitude and retried, recording the full impact assessment results; if the verification fails, fine-tuning feedback is provided, and the minimum executable rectification plan is regenerated; if the verification passes, version solidification and monitoring are integrated: the verified rectification items are solidified into versioned records, establishing a one-to-one mapping of "rule ID -> version -> log", and integrated with Sentinel monitoring for continuous tracking; if the version information is refreshed, the input is re-entered; the structured output includes: the final rectification plan & verification results, updated version records, and monitoring configuration.

[0130] In this embodiment, the complex credit approval chain analysis is effectively transformed into a series of specific and actionable rectification plans. The source tracing report serves as the core basis, ensuring the precise positioning of rectification actions and avoiding blind adjustments. The determination of the rules to be adjusted, the time points, and the breakpoints further refines the adjustment plans, enabling them to directly address the core of the problem rather than affecting the entire process. The final determined minimum adjustment entry point not only reduces rectification costs but also minimizes the impact on normal business processes, ensuring the stability and efficiency of the approval system.

[0131] Figure 8 This is a schematic diagram of the process for detecting and tracing implicit discrimination in credit approval based on counterfactual causal mapping and hierarchical SHAP value attribution according to Embodiment 1 of the present invention, as shown below. Figure 8 As shown, the process includes the following:

[0132] (1) S1 causal skeleton construction: abstract scores, rule blocks, protected attributes, proxy variables, manual review, and approval conclusions as nodes; label the direction of influence and propagation restrictions and map rule ID, threshold, and version.

[0133] (2) Causal scanning and path judgment: along the protected attribute / agent -> approval conclusion path, distinguish between the reasonable business path Γ 1egit and the illegal causal path Γillicit.

[0134] (3) Path iteration: determine whether the path is an illegal causal path.

[0135] (4) If so, perform path mapping: project the illegal path segment onto the rule block ID, threshold position, and policy version timeline, and output a labeled causal graph. The graph is now complete.

[0136] (5) If not, then the business is reasonable.

[0137] (6) S2 Same-risk sample pair location: Using the “rule block-breakpoint-version” ternary index, pair samples of the same risk but different batches are formed by matching the nearest samples in the neighborhood window.

[0138] (7) Once the pairing is complete, determine whether the difference can be explained by a reasonable business path.

[0139] (8) If yes, discard and continue to locate the same risk sample pair; if no (cannot be explained), perform candidate sample screening: remove the explainable sample pair and retain the candidate set C_pair.

[0140] (9) Based on the candidate samples, perform S3 counterfactual generation: replace the protected attribute / proxy value, deduce the descendant only along the illegal path, freeze the reasonable business nodes, and output the counterfactual sample pair C_cf.

[0141] (10) If counterfactual generation is performed, determine whether the window is consistent with the version. If not (parameter tuning and regeneration), perform counterfactual generation again. If yes, perform restricted hierarchical SHAP attribution.

[0142] (11) S4 Restricted Hierarchical SHAP Attribution: Construct a restricted alliance, calculate the marginal contribution vector Φ only within the scope of illegal descendant features and feature group -> rule block -> version.

[0143] (12) Based on the contribution vector Φ, perform S5 evidence package encapsulation: aggregate Φ to the rule level and generate evidence packages Epkg such as heatmap, sample pairs, path labels, and log fragments.

[0144] (13) Based on the source tracing report, carry out the S6 minimum rectification plan: locate the threshold breakpoint neighborhood, generate the rectification plan, and verify it in the sandbox.

[0145] (14) After the rectification is implemented, determine whether the verification is passed. If yes, solidify the version and monitor it with sentinels: update the causal graph, version record and log reference to complete the discovery -> rectification -> monitoring closed loop; if no, make fine adjustments to the threshold or range: adjust according to the minimum magnitude of the regression results, re-execute the verification, and re-verify the minimum rectification plan after adjustment.

[0146] The credit approval adjustment method based on counterfactual causal graphs provided in this application first abstracts the approval process modules into nodes, forming a directed causal skeleton and marking path attributes. Then, combining rule blocks, risk segments, and strategy version timelines, candidate samples with the same risk but different batches are selected from the case library. Next, based on abnormal causal paths, counterfactual comparisons are generated for each candidate sample, forming a compliance difference sample set. Subsequently, the approval chain is decomposed hierarchically according to feature groups, rule blocks, and conclusions. The contribution value of the hierarchical SHAP limited by abnormal paths is calculated using counterfactual samples, and rule-level impact details are output. Finally, the minimum adjustment entry point of the specific rule threshold is located based on the contribution vector. Through versioned recording and regression verification, accurate and traceable rectification of the credit approval chain is achieved, effectively improving the automation level and rectification efficiency of algorithm compliance governance, thereby achieving the technical effect of accurately locating and analyzing implicit discrimination.

[0147] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0148] Example 2

[0149] This application also provides a credit approval adjustment device based on counterfactual causal graphs. It should be noted that this counterfactual causal graph-based credit approval adjustment device can be used to execute the counterfactual causal graph-based credit approval adjustment method provided in this application. The following describes the counterfactual causal graph-based credit approval adjustment device provided in this application.

[0150] According to an embodiment of this application, an apparatus for implementing the above-described credit approval adjustment method based on counterfactual causal graphs is also provided. Figure 9 This is a schematic diagram of an optional credit approval adjustment device based on a counterfactual causal graph according to an embodiment of the present invention, such as... Figure 9 As shown, the credit approval adjustment device may include: a construction unit 90, a first determination unit 91, a generation unit 92, a second determination unit 93, and a third determination unit 94.

[0151] Among them, the construction unit 90 is used to construct a causal graph based on the current credit approval chain. The nodes in the causal graph are abstracted from the various process modules on the current credit approval chain. The directed edges in the causal graph are used to indicate the relationship between nodes. The causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths.

[0152] The first determining unit 91 is used to determine a set of candidate samples from the case library based on the set of rule blocks, risk segmentation set, strategy version timeline and causal graph corresponding to the current credit approval chain. The case library includes multiple credit case samples that have been approved.

[0153] The generation unit 92 is used to generate counterfactual samples corresponding to each candidate sample in the candidate sample set based on the abnormal causal path, to obtain a counterfactual sample pair set, wherein the counterfactual sample pairs in the counterfactual sample pair set are composed of candidate samples and counterfactual samples corresponding to the candidate samples;

[0154] The second determining unit 93 is used to split the current credit approval chain to obtain a set of feature groups, a set of rule blocks and an approval conclusion, and based on the counterfactual sample pair set, to determine the contribution value of each feature group in the feature group set under each rule block in the rule block set, and to obtain a hierarchical contribution vector.

[0155] The third determining unit 94 is used to determine the minimum adjustment entry based on the hierarchical contribution vector, and to adjust the current credit approval chain based on the minimum adjustment entry.

[0156] The credit approval adjustment device based on counterfactual causal graph provided in this application first abstracts the approval process module into nodes, forming a directed causal skeleton and marking path attributes. Then, combining rule blocks, risk segments, and strategy version timelines, it selects candidate samples of different batches with the same risk from the case library. Next, based on abnormal causal paths, it generates counterfactual comparisons for each candidate sample, forming a compliance difference sample set. Subsequently, the approval chain is decomposed hierarchically according to feature groups, rule blocks, and conclusions. The counterfactual samples are used to calculate the hierarchical SHAP contribution value limited by abnormal paths, outputting rule-level impact details. Finally, the minimum adjustment entry point of the specific rule threshold is located based on the contribution vector. Through versioned recording and regression verification, it achieves accurate and traceable rectification of the credit approval chain, effectively improving the automation level and rectification efficiency of algorithm compliance governance, thereby achieving the technical effect of accurately locating and analyzing implicit discrimination.

[0157] Optionally, the construction unit includes: a first abstraction module, used to abstract each process module on the current credit approval chain into nodes, wherein the nodes include at least: scoring nodes, rule block nodes, manual review nodes, approval conclusion nodes, preset attribute nodes, proxy variable nodes, and business intermediary nodes; a first determination module, used to determine directed edges based on the business relationships between nodes, and to construct a causal skeleton based on all nodes and all directed edges; a second determination module, used to determine the preset attribute nodes and proxy variable nodes as source nodes, and to determine the preset paths starting from the source nodes and pointing to the approval conclusion nodes based on the causal skeleton; a third determination module, used to determine the path labels of each preset path based on the business intermediary nodes, and to mark the path labels on the causal skeleton to obtain an initial causal graph; and a first mapping module, used to map the path labels on the initial causal graph to the rule block set and the strategy version timeline, to establish the correspondence between path identifiers, rule identifiers, and strategy version numbers, and to write the correspondence into the initial causal graph to obtain the causal graph.

[0158] Optionally, the first determining unit includes: a fourth determining module, used to map the causal graph to the rule block set, risk segment set, and strategy version timeline, and determine the breakpoint set for each rule block, wherein each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point; a first classification module, used to classify all credit case samples in the case library according to the rule blocks and segment them according to the strategy version timeline to obtain an index; a fifth determining module, used to determine the set of credit case samples located in the same risk segment and with the same credit information based on the index and the risk segment set, and to determine two adjacent credit case samples in the credit case sample set as sample pairs according to the time order; a first filtering module, used to filter all sample pairs based on the approval conclusion of the credit case samples to obtain an initial sample pair set; and a first processing module, used to process the initial sample pair set based on the causal graph to obtain an abnormal path descendant feature set and a candidate sample set.

[0159] Optionally, the generation unit includes: a sixth determining module, used to determine the set of node values ​​of the source node and freeze the reasonable business paths on the causal graph; a first replacement module, used to replace the original values ​​of the preset attributes or proxy variables indicated by the source node in each candidate sample with the node values ​​in the set of node values ​​to obtain the replaced candidate sample; a first derivation module, used to derive the replaced candidate sample on the frozen causal graph to obtain the initial counterfactual sample; and a first verification module, used to verify each initial counterfactual sample with the breakpoint set, the risk segmentation set, and the strategy version timeline to obtain the counterfactual sample and update the abnormal path descendant features of the abnormal causal paths affected by the counterfactual sample.

[0160] Optionally, the second determining unit includes: a first establishing module for establishing a mapping table between feature groups and rule blocks; a first comparison module for comparing each counterfactual sample pair in the counterfactual sample pair set with counterfactual samples as a benchmark to obtain normal features and freeze all normal features; a first adding module for adding all abnormal path descendant features to the current credit approval chain for testing to obtain test results, wherein the test results include at least: the abnormal causal paths affected by each abnormal path descendant feature; a first construction module for constructing a restricted test environment based on the test results; and a first calculation module for calculating the contribution value of each feature group under each rule block in the restricted test environment using a layered algorithm to obtain a layered contribution vector.

[0161] Optionally, the credit approval adjustment device further includes: a first alignment module, used to align the hierarchical contribution vector with the rule block and strategy version timeline after obtaining the hierarchical contribution vector to generate an evidence chain; a first generation module, used to generate a heat map based on the evidence chain; a first integration module, used to integrate the rule block and strategy version according to a preset order based on the hierarchical contribution vector to obtain an evidence report; and a first encapsulation module, used to encapsulate the evidence chain, heat map, and evidence report to obtain a source tracing report.

[0162] Optionally, the third determining unit includes: a seventh determining module, used to determine the rules to be adjusted and the time point of each rule to be adjusted based on the source tracing report; an eighth determining module, used to determine the breakpoint to be adjusted for each rule to be adjusted based on the time point of each rule to be adjusted; and a ninth determining module, used to determine the minimum adjustment entry point based on all rules to be adjusted, the time points of all rules to be adjusted, and the breakpoints to be adjusted for all rules to be adjusted.

[0163] The aforementioned credit approval adjustment device may also include a processor and a memory. The aforementioned construction unit 90, first determining unit 91, generation unit 92, second determining unit 93, third determining unit 94, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0164] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, the minimum adjustment entry point is determined based on the hierarchical contribution vector, and the current credit approval chain is adjusted based on this minimum adjustment entry point.

[0165] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0166] It should be noted that the aforementioned construction unit 90, first determining unit 91, generation unit 92, second determining unit 93, and third determining unit 94 correspond to steps S201 to S205 in Embodiment 1. The instances and application scenarios implemented by the aforementioned units and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the aforementioned units may be hardware or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The aforementioned units may also be part of a device and can run in the computer terminal 10 provided in Embodiment 1.

[0167] Example 3

[0168] Embodiments of this application may provide an electronic device. Figure 10 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 10 As shown, the electronic device may include: one or more ( Figure 10 (Only one is shown) processor 1002, memory 1004, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0169] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the credit approval adjustment method and apparatus based on counterfactual causal graphs in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned credit approval adjustment method based on counterfactual causal graphs. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0170] The processor can access information and applications stored in memory via a transmission device to execute the following steps: Based on the current credit approval chain, construct a causal graph, where nodes in the causal graph are abstracted from various process modules on the current credit approval chain, and directed edges in the causal graph indicate the relationships between nodes; the causal graph carries path labels, which indicate reasonable or abnormal causal paths in the business process; based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain, determine a candidate sample set from a case library, where the case library includes: multiple completed approvals... Credit case samples; based on abnormal causal paths, generate counterfactual samples corresponding to each candidate sample in the candidate sample set, resulting in a counterfactual sample pair set, where each counterfactual sample pair in the counterfactual sample pair set consists of a candidate sample and its corresponding counterfactual sample; decompose the current credit approval chain to obtain a feature group set, a rule block set, and an approval conclusion, and based on the counterfactual sample pair set, determine the contribution value of each feature group in the feature group set under each rule block in the rule block set, resulting in a hierarchical contribution vector; based on the hierarchical contribution vector, determine the minimum adjustment entry point, and adjust the current credit approval chain based on the minimum adjustment entry point.

[0171] The processor can invoke information and applications stored in memory via a transmission device to execute the following steps: Abstracting each process module in the current credit approval chain into nodes, where each node includes at least: a scoring node, a rule block node, a manual review node, an approval conclusion node, a preset attribute node, a proxy variable node, and a business intermediary node; determining directed edges based on the business relationships between nodes, and constructing a causal skeleton based on all nodes and all directed edges; identifying preset attribute nodes and proxy variable nodes as source nodes, and determining preset paths from the source nodes to the approval conclusion node based on the causal skeleton; determining path labels for each preset path based on the business intermediary nodes, and marking the path labels on the causal skeleton to obtain an initial causal graph; mapping the path labels on the initial causal graph to the rule block set and the strategy version timeline, establishing a correspondence between path identifiers, rule identifiers, and strategy version numbers, and writing the correspondence into the initial causal graph to obtain the causal graph.

[0172] The processor can access information and applications stored in memory via a transmission device to perform the following steps: mapping the causal graph to a set of rule blocks, a set of risk segments, and a timeline of policy versions to determine a set of breakpoints for each rule block, wherein each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point; classifying all credit case samples in the case library according to rule blocks and segmenting them according to the timeline of policy versions to obtain an index; based on the index and the set of risk segments, determining a set of credit case samples located in the same risk segment and with the same credit information, and determining two adjacent credit case samples in the credit case sample set as sample pairs according to time order; filtering all sample pairs based on the approval conclusions of the credit case samples to obtain an initial set of sample pairs; and processing the initial set of sample pairs based on the causal graph to obtain a set of abnormal path descendant features and a set of candidate samples.

[0173] The processor can access information and applications stored in memory via a transmission device to perform the following steps: determine the set of node values ​​for the source node and freeze the reasonable business paths on the causal graph; for each candidate sample, replace the original values ​​of preset attributes or proxy variables indicated by the source node in the candidate sample with the node values ​​from the set of node values ​​to obtain the replaced candidate sample; deduce the replaced candidate sample on the frozen causal graph to obtain the initial counterfactual sample; for each initial counterfactual sample, verify the initial counterfactual sample with the breakpoint set, risk segmentation set, and strategy version timeline to obtain the counterfactual sample, and update the abnormal path descendant features of the abnormal causal paths affected by the counterfactual sample.

[0174] The processor can access information and applications stored in memory via a transmission device to perform the following steps: establishing a mapping table between feature groups and rule blocks; comparing each counterfactual sample pair in the counterfactual sample pair set with counterfactual samples as a benchmark to obtain normal features, and freezing all normal features; adding all abnormal path descendant features to the current credit approval chain for testing, and obtaining test results, wherein the test results include at least: the abnormal causal paths affected by each abnormal path descendant feature; constructing a restricted test environment based on the test results; in the restricted test environment, using a layered algorithm to calculate the contribution value of each feature group under each rule block, and obtaining a layered contribution vector.

[0175] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: align the hierarchical contribution vector with the rule blocks and policy version timelines to generate an evidence chain; generate a heatmap based on the evidence chain; integrate the rule blocks and policy versions according to a preset order based on the hierarchical contribution vector to obtain an evidence report; and encapsulate the evidence chain, heatmap, and evidence report to obtain a source tracing report.

[0176] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: based on the source tracing report, determine the rules to be adjusted and the time point of each rule to be adjusted; based on the time point of each rule to be adjusted, determine the breakpoint to be adjusted for each rule to be adjusted; based on all rules to be adjusted, the time points of all rules to be adjusted, and the breakpoints to be adjusted for all rules to be adjusted, determine the minimum adjustment entry point.

[0177] This application provides a scheme for adjusting credit approval based on counterfactual causal graphs. First, the approval process modules are abstracted into nodes, forming a directed causal skeleton, and path attributes are labeled. Then, combining rule blocks, risk segments, and strategy version timelines, candidate samples of different batches with the same risk are selected from the case library. Next, based on abnormal causal paths, counterfactual comparisons are generated for each candidate sample, forming a compliance difference sample set. Subsequently, the approval chain is decomposed hierarchically according to feature groups, rule blocks, and conclusions. The counterfactual samples are used to calculate the hierarchical SHAP contribution value limited by abnormal paths, outputting rule-level impact details. Finally, the minimum adjustment entry point for specific rule thresholds is located based on the contribution vector. Through versioned recording and regression verification, accurate and traceable rectification of the credit approval chain is achieved, effectively improving the automation level and rectification efficiency of algorithmic compliance governance, thereby achieving the technical effect of accurately locating and analyzing implicit discrimination.

[0178] Those skilled in the art will understand that Figure 10The structure shown is for illustrative purposes only. Electronic devices can also be terminal devices such as smartphones, tablets, PDAs, and mobile internet devices (MIDs). Figure 10 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 10 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 10 The different configurations shown.

[0179] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0180] Example 4

[0181] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the credit approval adjustment method based on counterfactual causal graphs provided in Embodiment 1.

[0182] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0183] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing steps of a credit approval adjustment method based on a counterfactual causal graph.

[0184] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0185] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0186] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0187] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0188] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0189] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0190] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A credit approval adjustment method based on counterfactual causal graphs, characterized in that, include: Based on the current credit approval chain, a causal graph is constructed. The nodes in the causal graph are abstracted from the various process modules on the current credit approval chain. The directed edges in the causal graph are used to indicate the relationship between nodes. The causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths. Based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain, a candidate sample set is determined from the case library, wherein the case library includes: multiple credit case samples that have been approved; Based on the abnormal causal path, a counterfactual sample corresponding to each candidate sample in the candidate sample set is generated to obtain a counterfactual sample pair set, wherein the counterfactual sample pair in the counterfactual sample pair set is composed of the candidate sample and the counterfactual sample corresponding to the candidate sample; The current credit approval chain is split to obtain a feature group set, a rule block set, and an approval conclusion. Based on the counterfactual sample pair set, the contribution value of each feature group in the feature group set under each rule block in the rule block set is determined to obtain a hierarchical contribution vector. Based on the hierarchical contribution vector, the minimum adjustment entry point is determined, and the current credit approval chain is adjusted based on the minimum adjustment entry point.

2. The credit approval adjustment method according to claim 1, characterized in that, Based on the current credit approval chain, the steps to construct a causal graph include: The process modules on the current credit approval chain are abstracted as nodes, wherein the nodes include at least: scoring nodes, rule block nodes, manual review nodes, approval conclusion nodes, preset attribute nodes, proxy variable nodes, and business intermediary nodes; Based on the business relationships between the nodes, the directed edges are determined, and a causal skeleton is constructed based on all the nodes and all the directed edges. The preset attribute node and the proxy variable node are determined as source nodes, and based on the causal skeleton, a preset path is determined that starts from the source node and points to the approval conclusion node; Based on the business intermediary node, the path label of each preset path is determined, and the path label is marked on the causal skeleton to obtain an initial causal graph; The path labels on the initial causal graph are mapped to the rule block set and the policy version timeline to establish a correspondence between path identifiers, rule identifiers and policy version numbers, and the correspondence is written into the initial causal graph to obtain the causal graph.

3. The credit approval adjustment method according to claim 1, characterized in that, The steps for determining a candidate sample set from the case library based on the rule block set, risk segmentation set, strategy version timeline, and causal graph corresponding to the current credit approval chain include: The causal graph is mapped to the rule block set, the risk segment set, and the strategy version timeline to determine the breakpoint set for each rule block. Each breakpoint in the breakpoint set includes at least: a threshold, a segment boundary, and a version switching point. All credit case samples in the case library are categorized according to the rule blocks and segmented according to the strategy version timeline to obtain an index; Based on the index and the risk segment set, a set of credit case samples that are located in the same risk segment and have the same credit information is determined, and two adjacent credit case samples in the set of credit case samples are determined as sample pairs according to the time order. Based on the approval conclusions of the credit case samples, all the sample pairs are filtered to obtain an initial sample pair set; Based on the causal graph, the initial sample pair set is processed to obtain the abnormal path descendant feature set and the candidate sample set.

4. The credit approval adjustment method according to claim 1, characterized in that, Based on the aforementioned abnormal causal path The step of generating a counterfactual sample corresponding to each candidate sample in the candidate sample set includes: Determine the set of node values ​​for the source node and freeze the reasonable business paths on the causal graph; For each candidate sample, the original value of the preset attribute or proxy variable indicated by the source node in the candidate sample is replaced with the node value in the set of node values ​​to obtain the replaced candidate sample; The replaced candidate samples are used to deduce the initial counterfactual samples from the frozen causal graph. For each initial counterfactual sample, the initial counterfactual sample is checked against the breakpoint set, the risk segmentation set, and the strategy version timeline to obtain the counterfactual sample, and the abnormal path descendant features of the abnormal causal path affected by the counterfactual sample are updated.

5. The credit approval adjustment method according to claim 1, characterized in that, The step of determining the contribution value of each feature group in the feature group set under each rule block in the rule block set based on the counterfactual sample pair set, and obtaining the hierarchical contribution vector, includes: Establish a mapping table between the feature groups and the rule blocks; Using the counterfactual samples as a benchmark, each counterfactual sample pair in the counterfactual sample pair set is compared to obtain normal features, and all normal features are frozen. All abnormal path descendant features are added to the current credit approval chain for testing to obtain test results, wherein the test results include at least: the abnormal causal path affected by each of the abnormal path descendant features; Based on the test results, a restricted test environment is constructed; In the restricted testing environment, a layered algorithm is used to calculate the contribution value of each feature group under each rule block to obtain the layered contribution vector.

6. The credit approval adjustment method according to claim 1, characterized in that, After obtaining the hierarchical contribution vector, the following is also included: Align the hierarchical contribution vector with the rule block and the strategy version timeline to generate a chain of evidence; Based on the aforementioned chain of evidence, a heatmap is generated; Based on the hierarchical contribution vector, the rule blocks and policy versions are integrated according to a preset order to obtain an evidence report; The evidence chain, the heatmap, and the evidence report are encapsulated to obtain a source tracing report.

7. The credit approval adjustment method according to claim 6, characterized in that, The step of determining the minimum adjustment entry point based on the hierarchical contribution vector includes: Based on the source tracing report, determine the rules to be adjusted and the time point for each rule to be adjusted; Based on the time point of each rule to be adjusted, determine the breakpoint to be adjusted for each rule to be adjusted; Based on all the rules to be adjusted, the time points of all the rules to be adjusted, and the breakpoints of all the rules to be adjusted, the minimum adjustment entry point is determined.

8. A credit approval adjustment device based on counterfactual causal graphs, characterized in that, include: The construction unit is used to construct a causal graph based on the current credit approval chain. The nodes in the causal graph are abstracted from the various process modules on the current credit approval chain. The directed edges in the causal graph are used to indicate the relationship between nodes. The causal graph carries path labels, which are used to indicate reasonable business paths or abnormal causal paths. The first determining unit is used to determine a set of candidate samples from the case library based on the set of rule blocks, the set of risk segments, the timeline of strategy versions, and the causal graph corresponding to the current credit approval chain. The case library includes multiple credit case samples that have been approved. The generation unit is configured to generate a counterfactual sample corresponding to each candidate sample in the candidate sample set based on the abnormal causal path, thereby obtaining a counterfactual sample pair set, wherein the counterfactual sample pairs in the counterfactual sample pair set are composed of the candidate sample and the counterfactual sample corresponding to the candidate sample; The second determining unit is used to split the current credit approval chain to obtain a feature group set, a rule block set, and an approval conclusion, and based on the counterfactual sample pair set, to determine the contribution value of each feature group in the feature group set under each rule block in the rule block set, thereby obtaining a hierarchical contribution vector. The third determining unit is used to determine the minimum adjustment entry based on the hierarchical contribution vector, and to adjust the current credit approval chain based on the minimum adjustment entry.

9. A computer program product, characterized in that, The method includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the credit approval adjustment method based on counterfactual causal graphs as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the credit approval adjustment method based on counterfactual causal graphs as described in any one of claims 1 to 7.

Images