Authentication risk supervision decision-making method and device capable of explaining attribution and optimizing

Abstract

The invention provides an authentication risk supervision decision-making method and device capable of explaining attribution and optimization. The method comprises the steps of obtaining multi-dimensional feature vectors corresponding to basic information, qualification information, certificate issuing information, personnel information, supervision feedback information, authentication activity information and innovation activity information of an authentication subject; obtaining a risk value corresponding to the multi-dimensional feature vector through a risk prediction model; if the risk value is greater than an early warning threshold value, generating a Shapley value corresponding to each dimension in the multi-dimensional feature vector based on the feature vector and a risk prediction model by adopting a Shapley value formula; based on the feature vector corresponding to the target dimension, the risk prediction model and the constraint condition, performing anti-fact deduction by taking optimal modification as a target; and converting the optimal dimension and the disturbance vector corresponding to the optimal dimension into a natural language for indicating a rectification direction, and generating a risk rectification report. And an explainable and operable supervision closed loop from risk early warning to precise treatment is realized.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an interpretable attribution and optimization certification risk monitoring and decision-making method, an interpretable attribution and optimization certification risk monitoring and decision-making device, a computer device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Supervisory and management bodies refer to organizations that, within a management system, supervise, inspect, evaluate, guide, and, when necessary, impose sanctions on the compliance, quality, safety, or behavior of specific entities (such as enterprises, institutions, individuals, products, activities, etc.). Examples include administrative regulatory agencies and industry self-regulatory organizations. Certification bodies refer to professional institutions or organizations that, based on specific standards, specifications, and technical procedures, conduct conformity assessments of products, services, management systems, personnel, etc., and issue certification certificates to prove their conformity. Examples include quality certification centers. To prevent certification bodies from damaging the credibility of certification due to insufficient capabilities, irregular operations, or lack of credibility, thereby harming consumer rights and disrupting industry order, supervisory and management bodies are needed to conduct risk supervision of certification bodies.

[0003] Currently, supervisory bodies primarily employ machine learning-based risk prediction models (such as random forests, gradient boosting decision trees, or deep neural networks) to conduct risk supervision of certified entities. Specifically, this involves first collecting historical data from multiple certified entities and labeling each entity with a risk value. This historical data and risk values ​​are then used as a training set. Next, the machine learning model is trained using this training set. Finally, for the certified entity to be supervised, relevant data is obtained and input into the trained machine learning model. The model outputs a specific numerical value, which represents the risk value of the certified entity to be supervised. Supervisory bodies then use this risk value to supervise the certified entity.

[0004] However, machine learning models, acting as black boxes, only provide risk values ​​and cannot reveal the specific causes of risks or the attribution of responsibility. This leads to a lack of interpretability and persuasiveness in regulatory decisions, resulting in regulatory problems stemming from opaque legal basis. Furthermore, machine learning models only provide risk warnings and cannot generate actionable rectification guidelines based on attribution results, making it difficult for supervisory and regulatory bodies to guide certification bodies in effectively reducing risks. Therefore, there is an urgent need to provide an interpretable and actionable regulatory decision-making method to penetrate the risk black box and achieve a paradigm shift from early warning to governance. Summary of the Invention

[0005] The purpose of this application is to provide an interpretable, attributable, and optimized certification risk monitoring and decision-making method, device, computer equipment, computer-readable storage medium, and computer program product, so as to provide an interpretable and operable regulatory decision report and realize the paradigm upgrade of certification supervision from early warning to governance.

[0006] To address the aforementioned technical problems, this application provides the following technical solutions: The first aspect of this application provides an interpretable attribution and optimization-based certification risk monitoring decision-making method. The method includes: acquiring multi-dimensional feature vectors corresponding to the certification entity's basic information, qualification information, certificate issuance information, personnel information, monitoring feedback information, certification activity information, and innovation activity information; obtaining risk values ​​corresponding to the multi-dimensional feature vectors through a risk prediction model, wherein the risk prediction model is a pre-trained machine learning model capable of generating risk values ​​for the certification entity based on its basic information, qualification information, certificate issuance information, personnel information, monitoring feedback information, certification activity information, and innovation activity information; if the risk value exceeds a warning threshold, the Shapley value formula is used. Based on the feature vector and risk prediction model, a Shapley value is generated for each dimension in the multidimensional feature vector, and one or more target dimensions to be adjusted are selected based on the Shapley value. Based on the feature vector corresponding to the target dimension, the risk prediction model, and the constraints, counterfactual inference is performed with the goal of optimal modification to obtain the optimal dimension and its corresponding perturbation vector in the target dimension. The constraints are that the feature vector of the dimension with immutable information remains unchanged and the perturbation vector conforms to the business logic. The optimal dimension and its corresponding perturbation vector are converted into natural language to indicate the direction of rectification, and a risk rectification report including risk value, dimensions with positive Shapley values, and natural language is generated.

[0007] Compared to existing technologies, the interpretable attribution and optimization-based certification risk supervision decision-making method provided in the first aspect of this application uses Shapley values ​​to perform fair and transparent risk attribution on machine learning models, clarifying the marginal contribution of each feature to risk. This makes regulatory decisions interpretable and persuasive. Furthermore, through constraint optimization and counterfactual deduction, it automatically generates the lowest-cost and most effective rectification path while ensuring the feasibility of business logic, and translates the results into natural language instructions to guide certification entities in implementing targeted rectification. This achieves an interpretable and operable regulatory closed loop from risk warning to precise governance, significantly improving the accuracy, compliance, and execution efficiency of supervision.

[0008] In other embodiments provided in this application, based on the feature vector, risk prediction model and constraints corresponding to the target dimension, counterfactual deduction is performed with the goal of optimal modification to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. This includes: based on the feature vector, risk prediction model, constraints and cost weights corresponding to the target dimension, counterfactual deduction is performed with the goal of minimizing the total cost of optimal modification to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The cost weights are used to indicate the cost required for modification of each dimension.

[0009] By introducing cost weights and explicitly using counterfactual reasoning with the goal of minimizing total cost, the operability and economy of the rectification recommendations are further enhanced. It not only focuses on whether the risk has been reduced to the safety threshold, but also quantifies the input costs of different rectification measures from the perspective of management practice. This ensures that the final optimized path meets regulatory requirements while minimizing the implementation burden on the certification body, thus upgrading the regulatory decision from theoretically feasible to practically optimal. This improves the acceptability and implementation efficiency of the rectification plan and achieves a scientific balance between regulatory compliance and institutional operating costs.

[0010] In other embodiments provided in this application, based on the feature vectors, risk prediction models, constraints, and cost weights corresponding to the target dimension, and with the goal of minimizing the total cost of the optimal modification, counterfactual deduction is performed to obtain the optimal dimension and its corresponding perturbation vector within the target dimension. This includes: constructing an optimization path graph based on the feature vectors, risk prediction models, constraints, and cost weights corresponding to the target dimension, wherein each coordinate axis of the optimization path graph represents a different dimension, corresponding constraints are set on the corresponding coordinate axes, points in the optimization path graph indicate risk values ​​under different feature vectors, and the optimization path graph also includes multiple risk contour lines, with multiple lines configured from the current risk value to the warning threshold; outputting the optimization path graph; receiving the target line selected based on the optimization path graph, and calculating the optimal dimension and its corresponding perturbation vector within the target dimension based on the target line.

[0011] By introducing a visual optimization path diagram, the abstract optimization process is transformed into an intuitive interactive interface, enabling intelligent decision support for humans on the loop. Regulatory personnel can clearly identify the costs and effects of different rectification paths in the diagram and autonomously select the optimal route based on regulatory preferences and actual circumstances. This enhances the flexibility of the method and the human-machine collaboration capability, upgrading risk governance from fully automated deduction to visual assisted decision-making, and improving the adaptability of regulatory responses.

[0012] In other embodiments provided in this application, selecting one or more target dimensions to be adjusted based on Shapley values ​​includes: sorting multiple dimensions in descending order of Shapley values; and selecting a preset number of dimensions with positive Shapley values ​​at the top of the sort as target dimensions.

[0013] By sorting by Shapley value and filtering positive contribution dimensions, the priority of risk causes was focused, optimizing the rectification target from full-dimensional coverage to key point entry. The core indicators with the most significant impact on risk were automatically identified, guiding regulatory resources and rectification measures to be precisely directed towards the most critical risk drivers. This effectively avoided the dispersion and generalization of rectification directions, significantly improved the efficiency and pertinence of risk governance, and reduced unnecessary management costs.

[0014] In other embodiments provided in this application, selecting a preset number of dimensions with positive Shapley values ​​at the top of the sorted order as target dimensions includes: constructing a waterfall plot based on the Shapley values ​​corresponding to the feature vectors of each dimension according to the sorting of multiple dimensions, wherein the horizontal axis of the waterfall plot represents the Shapley value and the vertical axis represents each dimension; outputting the waterfall plot; receiving multiple dimensions selected based on the waterfall plot, and determining the selected multiple dimensions as target dimensions.

[0015] By visually displaying the magnitude and sign of Shapley values ​​across various dimensions using a waterfall chart, complex attribution results are transformed into a visualized risk structure diagram. This allows regulators to flexibly select remediation targets based on the risk contribution distribution on the chart, combined with domain knowledge. This enhances the transparency of the methodology and the interactivity of decision-making, shifting risk attribution from a black-box output to a white-box collaboration, thereby improving the intuitiveness of regulatory explanations and the acceptability of decisions.

[0016] In other embodiments provided in this application, before obtaining the risk value corresponding to the multidimensional feature vector through the risk prediction model, the method further includes: predicting the future feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information of the certification subject after a preset time interval based on the multidimensional feature vectors; obtaining the risk value corresponding to the multidimensional feature vectors through the risk prediction model includes: obtaining the risk value corresponding to the future feature vectors through the risk prediction model.

[0017] By introducing a prediction mechanism based on future feature vectors, risk supervision has been upgraded from static assessment to dynamic early warning. It can simulate the potential state of certified entities in the future based on current data and identify their risk evolution trends in advance. This transforms regulatory intervention from post-event rectification to pre-event prevention, significantly enhancing the initiative and timeliness of supervision. It helps to implement precise guidance before risks become apparent and prevent the occurrence of systemic risks from the source.

[0018] In other embodiments provided in this application, there are multiple optimal dimensions. Before generating a risk remediation report including risk values, positive Shapley values, and natural language, the method further includes: determining the time cost of each optimal dimension, the dependencies between each optimal dimension during execution, and the target dimension among the multiple optimal dimensions that can achieve emergency risk mitigation; sorting each optimal dimension in ascending order of time cost; adjusting dependent dimensions to adjacent dimensions from back to front in the sorted optimal dimensions; placing the target dimension at the front in the adjusted optimal dimensions; generating time suggestion information for completion within a positive correlation time period for the currently sorted optimal dimensions; generating a risk remediation report including risk values, positive Shapley values, natural language, and time suggestion information, including: generating a risk remediation report including risk values, positive Shapley values, natural language, and time suggestion information.

[0019] By comprehensively considering time costs, execution dependencies, and urgent mitigation goals, the system intelligently prioritizes and schedules multiple optimal dimensions to construct a structured, phased rectification roadmap. It not only provides static suggestions on what and how much to change, but also clarifies dynamic execution strategies regarding when and in what order to change. This makes the rectification plan more aligned with actual management processes and resource constraints, significantly improving the operability, timing, and execution consistency of the rectification scheme, and facilitating efficient and orderly risk mitigation.

[0020] The second aspect of this application provides an interpretable attribution and optimization-based certification risk monitoring and decision-making device. The device includes: a vector acquisition module for acquiring multi-dimensional feature vectors corresponding to the certification entity's basic information, qualification information, certificate issuance information, personnel information, monitoring feedback information, certification activity information, and innovation activity information; a risk prediction module for obtaining risk values ​​corresponding to the multi-dimensional feature vectors through a risk prediction model, wherein the risk prediction model is a pre-trained machine learning model capable of generating risk values ​​for the certification entity based on its basic information, qualification information, certificate issuance information, personnel information, monitoring feedback information, certification activity information, and innovation activity information; and an attribution interpretation module for using Shapley if the risk value exceeds a warning threshold. The system comprises four modules: a value formula, which generates Shapley values ​​for each dimension of a multidimensional feature vector based on feature vectors and a risk prediction model, and selects one or more target dimensions to be adjusted based on the Shapley values; an optimal deduction module, which performs counterfactual deduction based on the feature vectors corresponding to the target dimensions, the risk prediction model, and constraints, with the goal of optimal modification, to obtain the optimal dimension and its corresponding perturbation vector, where the constraints are that the feature vectors of dimensions with immutable information remain unchanged and the perturbation vector conforms to business logic; and a decision report module, which converts the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, generating a risk rectification report including risk values, dimensions with positive Shapley values, and natural language.

[0021] A third aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method of the first aspect.

[0022] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of the first aspect.

[0023] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the method of the first aspect.

[0024] The interpretable attribution and optimization certification risk monitoring and decision-making device provided in the second aspect of this application, the computer equipment provided in the third aspect, the computer-readable storage medium provided in the fourth aspect, and the computer program product provided in the fifth aspect have the same or similar beneficial effects as the interpretable attribution and optimization certification risk monitoring and decision-making method provided in the first aspect. Attached Figure Description

[0025] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein: Figure 1 This is a schematic diagram illustrating an application scenario for the certification risk monitoring and decision-making method that can be explained and optimized according to the embodiments of this application. Figure 2 This is a flowchart illustrating the certification risk monitoring and decision-making method that allows for attribution and optimization in the embodiments of this application. Figure 1 ; Figure 3 This is a flowchart illustrating the certification risk monitoring and decision-making method that allows for attribution and optimization in the embodiments of this application. Figure 2 ; Figure 4 This is a schematic diagram of a waterfall diagram in an embodiment of this application; Figure 5 This is a schematic diagram of the optimization path graph in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the certification risk monitoring and decision-making device that can explain attribution and optimization in the embodiments of this application. Figure 1 ; Figure 7 This is a schematic diagram of the structure of the certification risk monitoring and decision-making device that can explain attribution and optimization in the embodiments of this application. Figure 2 ; Figure 8 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0026] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0027] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0028] Currently, risk supervision methods based on machine learning lack interpretability, cannot reveal the specific causes and attributions of risks, and can only provide risk warnings without offering actionable rectification guidelines, resulting in opaque regulatory decision-making and insufficient governance effectiveness.

[0029] In view of this, embodiments of this application provide an interpretable attribution and optimization certification risk supervision decision-making method, an interpretable attribution and optimization certification risk supervision decision-making device, a computer device, a computer-readable storage medium, and a computer program product. Based on traditional machine learning risk early warning, it introduces the Shapley value algorithm to transparently attribute risk prediction models, accurately locate key risk drivers, and then, combined with business constraints and cost optimization objectives, search for the lowest-cost rectification path that meets regulatory requirements in the feature space through counterfactual deduction. Finally, the mathematical results are transformed into executable natural language rectification instructions, thereby realizing a new paradigm of intelligent supervision with interpretable risks, quantifiable attributions, operable rectification, and closed-loop decision-making.

[0030] It should be noted that all components, data, and related processing methods involved in this application are authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0031] First, the application scenarios of the interpretable attribution and optimization certification risk monitoring decision-making method provided in the embodiments of this application will be described.

[0032] Figure 1 This is a schematic diagram illustrating an application scenario for the certification risk monitoring and decision-making method that allows for attribution and optimization, as described in this application. (See also...) Figure 1 As shown, the scenario may include: risk prediction model 11, Shapley value formula module 12, and counterfactual deduction module 13.

[0033] Risk prediction model 11 is a machine learning model that automatically identifies risk patterns and outputs quantified risk values ​​based on the multi-dimensional feature data of the certified entity. This model establishes a mapping function from the feature space to the risk score by learning the complex nonlinear relationship between features and risk labels in a large amount of historical regulatory data.

[0034] In practical applications, risk prediction model 11 can adopt gradient boosting decision tree-based models (such as XGBoost, LightGBM, CatBoost), ensemble learning models (such as random forests), deep learning models (such as fully connected neural networks), and traditional classification models such as support vector machines and logistic regression.

[0035] Shapley value formula module 12 contains the Shapley value formula. The Shapley value is a mathematical method derived from cooperative game theory, used to fairly evaluate the contribution of each participant to the total payoff of the alliance. In this application, each dimension of the feature vector of the certifying subject is considered a participant in the game, and the risk prediction value is considered the total payoff. The Shapley value is used to quantify the marginal contribution of each feature to the final risk value, ensuring that the attribution process conforms to fairness, efficiency, and additivity.

[0036] The formula for Shapley value is as follows: in, N Represents the set of all features. S Indicates that it does not contain features i Any feature subset, f ( S ) indicates the use of a subset. S The output value of the risk prediction model is a feature in the model. S | represents a subset S The number of features, ϕ i ( v This is the characteristic. i The Shapley value reflects its average marginal contribution.

[0037] This formula calculates the features by iterating through all possible combinations of features. i By incorporating the average increase or decrease in the prediction results for each combination, the risk contribution can be fully allocated and explained while maintaining the consistency of model predictions.

[0038] Counterfactual deduction module 13 includes a counterfactual deduction algorithm. In this application, counterfactual deduction refers to constructing a hypothetical new feature vector by systematically modifying the variable dimensions of the current feature vector while maintaining certain realistic constraints. This new feature vector reduces the predicted risk value below a safe threshold, minimizing the total cost of the modification. It answers the key regulatory question of what the minimum changes an organization needs to make to achieve compliance.

[0039] The objective function for counterfactual inference is: Among them, the first item The weighted Manhattan distance represents the total management cost of the rectification. The goal is to minimize this cost. The second term is f(x+δ)-T. safe This is a penalty item. Ensure that the post-rectification risk value f(x+δ) is lower than the safety threshold T. safeThe third item, C, is a penalty coefficient used to balance cost and compliance.

[0040] Apply hard constraints: for ∀j∈I immutable Forced δ j =0 (history cannot be altered). For numerical features, LB is satisfied. k ≤x k +δ k ≤UB k (For example, the number of personnel cannot be negative, nor can it exceed the industry's theoretical upper limit).

[0041] Since the objective function is non-convex (determined by the nonlinearity of f) and non-smooth (L1 norm), this embodiment uses a fast iterative shrinkage threshold algorithm or a projective gradient descent method to solve for the optimal perturbation vector δ. * .

[0042] Since this problem is a constrained nonlinear and nonconvex optimization problem, numerical methods such as the fast iterative threshold shrinkage algorithm and the projective gradient descent method can be used to solve it and obtain the optimal perturbation vector. δ ∗ This is then mapped into specific and actionable rectification instructions.

[0043] In actual operation, the system first acquires raw data such as the certification entity's basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information, and constructs them into a multi-dimensional feature vector. This vector is input into the risk prediction model 11, which calculates and outputs the corresponding risk value. If the risk value exceeds the warning threshold, the feature vector and the risk prediction model 11 are input into the Shapley value formula module 12. This module calculates the Shapley value corresponding to each feature and selects the target dimensions to be adjusted. Then, the feature vector of the target dimension, along with the risk prediction model 11, constraints, and cost weights, are input into the counterfactual inference module 13. This module uses an optimization algorithm to solve for the perturbation vector with the lowest cost. Finally, the perturbation vector is parsed and converted into rectification suggestions in natural language form, which, together with the risk value and key risk dimensions, generate a complete risk rectification report.

[0044] Next, the certification risk monitoring and decision-making method with explainable attribution and optimization provided in the embodiments of this application will be described in detail.

[0045] Figure 2 This is a flowchart illustrating the certification risk monitoring and decision-making method that allows for attribution and optimization in the embodiments of this application. Figure 1 See Figure 2 As shown, the method may include: S21: Obtain the multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification entity.

[0046] Basic information refers to the fundamental identity and status data generated by the certified entity during registration and daily operations. Examples include: organization name, unified social credit code, registered address, establishment date, legal representative, registered capital, organization type, and operating status.

[0047] Qualification information refers to the legally or industry-recognized qualifications and licenses that a certification body must possess to conduct certification activities. Examples include: the certification fields and scope approved by the Certification and Accreditation Administration of the People's Republic of China (CNCA), accreditation certificate numbers and validity periods, registration status of professional and technical personnel qualifications, and laboratory accreditation qualifications.

[0048] Certificate issuance information refers to the statistical and characteristic data of various certification certificates issued by the certification body to the certified entities during the certification business activities. Examples include: cumulative number of certificates issued, certificate category distribution (e.g., ISO 9001, ISO 14001), number of certificates within their validity period, number of suspended / revoked certificates, and certificate growth rate.

[0049] Personnel information refers to the configuration and status data of professional and technical personnel and management personnel directly related to certification activities within the certification entity. Examples include: the number of full-time auditors, technical experts, and management personnel; the personnel qualification structure (e.g., the ratio of senior to intermediate auditors); and the continuing education completion rate.

[0050] Supervisory feedback information refers to the evaluations, complaints, and handling records of certification entities' professional conduct and results from regulatory authorities, clients, or the general public. Examples include: the number of regulatory inspections, administrative penalty records, client complaint rates and processing times, certification result sampling pass rates, and social reputation ratings.

[0051] Certification activity information refers to the business process and behavioral data generated by the certification entity in carrying out specific certification services. Examples include: the number of audit projects per year, auditor days, timeliness of certification decisions, proportion of audits conducted in different locations, compliance of certification fees, and records of deviations in the implementation of audit plans.

[0052] Information on innovation activities refers to data on the innovation investment and practices of the certification entity in areas such as technology research and development, service models, and standard application. Examples include: R&D expenditure amount, number of national / industry standards participated in, number of patents or software copyrights granted, number of new technology and method application projects, and investment in digital transformation.

[0053] Generate multidimensional feature vectors x ={ x 1, x 2,…, x d}∈ R d First, the specific indicators under each type of information (such as the number of full-time auditors, certificate growth rate, complaint rate, etc.) are defined, quantified, and standardized. Then, all the quantified indicators are spliced ​​together in a predetermined order to form a numerical feature vector of a unified dimension, which can be input into the subsequent risk prediction model for calculation.

[0054] S22: Obtain the risk value corresponding to the multidimensional feature vector through the risk prediction model.

[0055] Among them, the risk prediction model is a pre-trained machine learning model that can generate the risk value of the certification entity based on the certification entity's basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information.

[0056] The risk prediction model has been explained in detail previously and will not be repeated here.

[0057] After obtaining the multidimensional feature vector x, it is input into a pre-trained risk prediction model f. This model is typically trained based on historical regulatory data (including feature vectors and corresponding manually labeled risk tags) and is capable of learning the complex nonlinear relationship between features and risks. After receiving the input vector x, the model performs forward inference through its internal algorithms (such as forward propagation in gradient boosting decision trees or hierarchical computation in neural networks), ultimately outputting a continuous probability value. P current= f ( x This value represents the current risk probability of the certified entity, and is usually in the range of [0,1]. It is used for subsequent threshold judgment and attribution analysis.

[0058] S23: If the risk value is greater than the warning threshold, the Shapley value formula is used to generate the Shapley value corresponding to each dimension in the multidimensional feature vector based on the feature vector and the risk prediction model, and one or more target dimensions to be adjusted are selected based on the Shapley value.

[0059] When risk value Pcurrent= f ( x When the risk level exceeds the preset warning threshold, it means that the current status of the certified entity, as quantitatively assessed by the model, has exceeded the level of safety or acceptable risk recognized by regulators, triggering the system's risk warning mechanism. This indicates that the entity has a high probability of potential problems in its operations, compliance, or quality, requiring regulatory intervention. Therefore, the system will automatically initiate subsequent explainable attribution analysis (Shapley value calculation) and rectification path optimization (counterfactual deduction) processes, aiming to shift from risk warning to root cause diagnosis and governance guidance, providing clear evidence and feasible action plans for regulatory decisions.

[0060] When calculating the Shapley value, firstly, the risk prediction model... f It is regarded as the rule of cooperative game, and all dimensions of the multidimensional feature vector are considered. N ={1,2,…, d They are considered as participants in the game. Next, for each feature dimension to be evaluated... i The algorithm will iterate through all cases that do not contain i Feature subset S ⊆ N ∖{ i For each subset S Calculate the feature subsets respectively S Model output value f ( S ) and added features i The output value after f ( S ∪{ i The difference between the two is [}). f ( S ∪{ i})− f ( S [)] is the characteristic i The marginal contribution within that specific subset. Finally, the marginal contributions across all possible subsets are distributed according to the subset size. A weighted average is then performed to calculate the final feature dimension. i Shapley value ϕ i This allows for the precise quantification of its average contribution to the final risk prediction value.

[0061] When selecting a specific target dimension, you can first sort all dimensions from highest to lowest according to their Shapley values, and then filter out all ϕ values. iDimensions with a value greater than 0 (i.e., dimensions that positively contribute to increasing risk). In automatic mode, a preset number of dimensions (e.g., the top 3) ranked by default can be selected as target dimensions to achieve focused intervention. In interactive or manual mode, regulators can select specific dimensions from all positively contributing dimensions, or select all positively contributing dimensions as targets, based on a ranked list or visualization (e.g., a waterfall chart), combined with domain knowledge. This allows subsequent optimization calculations to focus on these selected dimensions, ensuring flexibility and targeting in the rectification direction. Alternatively, counterfactual analysis can be performed using all dimensions to systematically explore the globally optimal solution, avoiding the omission of potential, cross-feature-linked, low-cost rectification paths due to local dimension selection.

[0062] If the risk value P current= f ( x If the risk level is less than or equal to the preset warning threshold, it means that the current operational status of the certified entity, as assessed by the model, is within the safe or acceptable risk range as determined by regulators, and no high-risk warning has been triggered. In this case, there is no need to initiate subsequent complex attribution analysis and rectification simulation processes. A brief report containing the current risk value and low-risk conclusion can be output, and the entity will be marked as a regular monitoring target. The process then ends or enters the next round of periodic assessment.

[0063] S24: Based on the feature vectors, risk prediction models and constraints corresponding to the target dimension, counterfactual deduction is performed with the goal of optimal modification to obtain the optimal dimension in the target dimension and its corresponding perturbation vector.

[0064] The constraints here are that the dimensional feature vectors of immutable information remain unchanged and the perturbation vectors conform to business logic. That is, the constraints here are used to ensure that the rectification plan generated by counterfactual reasoning is mathematically feasible and business-executable. They mainly include two types: the first type is immutable constraints, that is, constraints that apply to sets of immutable features. I For immutable dimensions (such as the number of historical administrative penalties, the date of establishment of the institution, etc.), force their perturbation values. δj =0, ensuring that historical facts are not altered. The second type is business logic constraints, which set feasible upper and lower bounds for variable features. LBk ≤ xk + δk ≤ UBk (For example, the number of full-time auditors cannot be negative after adjustments, and the monthly certificate issuance growth rate cannot exceed the industry experience limit), ensuring that the derived new feature vectors are accurate. x It falls within the reasonable business feasible domain Dfeasible.

[0065] In the specific deduction, in the feature space, the current feature vector corresponding to the target dimension is...x target, risk prediction model f Immutability, business logic constraints, and security thresholds T The concept of "safe" together constitutes a constrained optimization problem. Its core objective is to find an optimal perturbation vector. δ * This allows for a change in the risk value after rectification, provided only the target dimension is changed. f ( x target+ δ *) Below the safety threshold. This is because the objective function contains nonlinearity. f Given the non-smooth L1 norm, a numerical optimization algorithm (such as projective gradient descent) is used for iterative solution. The algorithm searches within the feasible region defined by the constraints and ultimately outputs the algorithm that minimizes the objective function. δ *, its non-zero components correspond to the optimal dimension (such as the number of full-time auditors) and the specific amount of change (such as... δ Personnel = +2).

[0066] S25: Convert the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, and generate a risk rectification report including risk value, dimension with positive Shapley value, and natural language.

[0067] When generating the report, first analyze the optimal perturbation vector. δ The non-zero elements in * convert each non-zero dimension and its corresponding specific perturbation value into structured natural language instructions through a predefined semantic mapping rule base. For example: if δ If personnel = +2, it maps to "It is recommended to add 2 full-time auditors with the appropriate qualifications". δ If the issuance rate is -0.15, it maps to "It is recommended to reduce the monthly issuance growth rate by 15 percentage points." Then, these instructions and initial risk values ​​are... P The system integrates the current value and all dimensions with positive Shapley values, along with their contributions, to automatically generate a structured report titled "Risk Remediation Recommendations." This report clearly identifies the risk level, key risk causes, and the mathematically lowest-cost compliance path.

[0068] Risk rectification reports can be issued to market supervision departments, which are the main bodies responsible for supervision and management, or to certification entities, which are the subjects of supervision.

[0069] Upon receiving the report, regulatory authorities can use it as a scientific basis for administrative enforcement or guidance. The report clarifies the risk attribution, making penalties or rectification requirements verifiable. Simultaneously, the report provides the most cost-effective rectification path, allowing regulatory authorities to issue specific and quantifiable rectification instructions and track and verify the rectification effects.

[0070] Upon receiving the report, the certified entity can regard it as a precise medical examination report and prescription. The report clearly identifies specific shortcomings (risk dimensions) in its operations and provides clear directions for improvement (natural language instructions). The organization's management can use this report to formulate internal rectification plans and allocate resources to reduce risks in the most efficient and compliant manner, thereby proactively meeting regulatory requirements.

[0071] As described above, the interpretable attribution and optimization-based certification risk supervision decision-making method provided in this application uses Shapley values ​​to perform fair and transparent risk attribution on machine learning models, clarifying the marginal contribution of each feature to risk. This makes regulatory decisions interpretable and persuasive. Furthermore, through constraint optimization and counterfactual deduction, it automatically generates the lowest-cost and most effective rectification path while ensuring the feasibility of business logic. The results are then converted into natural language instructions to guide certification entities in implementing targeted rectification. This achieves an interpretable and operable regulatory closed loop from risk warning to precise governance, significantly improving the accuracy, compliance, and execution efficiency of supervision.

[0072] Furthermore, as a response to Figure 2 In a refinement and extension of the method shown, this application embodiment also provides an interpretable attribution and optimization-based certification risk monitoring decision-making method.

[0073] Figure 3 This is a flowchart illustrating the certification risk monitoring and decision-making method that allows for attribution and optimization in the embodiments of this application. Figure 2 See Figure 3 As shown, the method may include: S31: Obtain the multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification entity.

[0074] In order to identify potential risks to certified entities in advance, we predict their future state based on current data and conduct a risk assessment of that future state.

[0075] S32: Based on multi-dimensional feature vectors, predict the future feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification subject after a preset time interval.

[0076] In specific forecasting, time series analysis or supervised learning models can be used. The current multidimensional feature vector x, along with historical time series data (if available), is taken as input. For each type of information (such as the number of certificates issued, personnel changes, etc.), a specific forecasting model (such as the ARIMA model for predicting numerical trends, or using recurrent neural networks such as LSTM to capture time-series dependencies) or inference based on business rules (such as extrapolating from the current growth rate) may be used to calculate its predicted value at a preset time interval (such as 6 months later). Finally, these predicted values ​​are reassembled to form the corresponding future feature vector x. future This serves as the basis for subsequent dynamic risk assessments.

[0077] For example, let's take a certification center as an example. Based on its current data (such as a monthly certificate issuance growth rate of 10% and 20 full-time auditors), we can make predictions. Using a time series model, we predict that its monthly certificate issuance growth rate may rise to 15% in six months (certification activity information). Simultaneously, based on its recruitment plan, we estimate that the number of full-time auditors will increase to 25 (personnel information). Combining these predicted values ​​with other extrapolated values ​​from other information (such as the complaint rate in monitoring feedback potentially increasing with business volume), we construct a future feature vector x describing its future operational status. future This allows for the assessment of potential future risks.

[0078] S33: Obtain the risk value corresponding to the future feature vector through a risk prediction model.

[0079] Compared to step S22 above, which uses the current feature vector x to calculate the current static risk value P current The key difference in step S33 here is that its input is the predicted future feature vector x. future This step will x future Input the same pre-trained risk prediction model f to calculate the future risk value P. future =f(x future This value represents the probability of potential risks for the certified entity at a preset future time point (such as 6 months later), thereby expanding the model's assessment capabilities from diagnosing the current situation to predicting future risk trends, achieving an upgrade from static assessment to dynamic early warning.

[0080] S34: If the risk value is greater than the warning threshold, the Shapley value formula is used to generate the Shapley value corresponding to each dimension in the multidimensional feature vector based on the feature vector and the risk prediction model.

[0081] S35: Sort multiple dimensions in descending order of Shapley value; select the first-ranked dimension with a preset number of positive Shapley values ​​as the target dimension.

[0082] After calculating the Shapley values ​​for all dimensions ϕi Then, first based on ϕi The values ​​are sorted in descending order to generate an ordered list. Then, the list is automatically iterated through, and all values ​​are selected. ϕi Dimensions >0 (i.e., dimensions that positively contribute to increasing current risk). Then, based on preset intervention depth parameters, i.e., preset quantities (e.g. K =3), from the list of positive contribution dimensions that have been filtered out, select the top K dimensions and directly determine them as the target dimensions to be adjusted. This process is fully automated, ensuring that subsequent counterfactual analysis can accurately focus on the most critical risk drivers.

[0083] The preset quantity here can be a fixed empirical value pre-set by regulatory experts based on historical cases and governance capabilities (e.g., always focusing on the top 3 main dimensions). Alternatively, it can be dynamically calculated based on the distribution of the attribution results, for example, by setting a cumulative contribution threshold (e.g., selecting a number of dimensions so that the sum of their Shapley values ​​exceeds 80% of the total positive contribution), and automatically determining the K value. Alternatively, this parameter can be designed as an interactive configurable item, allowing regulators to manually adjust it after reviewing the waterfall chart, thereby balancing the focus and comprehensiveness of the rectification efforts.

[0084] To visualize attribution results and incorporate human decision-making interaction, a waterfall chart can be used to visually display the risk contribution of each dimension and allow regulators to manually select target dimensions in the chart based on their professional judgment.

[0085] Specifically, step S35 above may include: S351: Based on the Shapley values ​​corresponding to the feature vectors of each dimension, a waterfall plot is constructed according to the sorting of multiple dimensions. The horizontal axis of the waterfall plot represents the Shapley value, and the vertical axis represents each dimension.

[0086] In the specific construction, first list the entries for each dimension, already sorted in descending order of Shapley values, as the ordinate. When plotting, use the zero point of the x-axis as a reference, and accumulate the Shapley values ​​for each dimension from left to right (or from bottom to top). The bar for the first dimension starts at the zero point and extends to a length equal to its zero point. ϕ 1. The second dimension bar begins at the end of the previous bar and extends by a length equal to... ϕ 2. And so on, until finally, all the bars with positive Shapley values ​​are summed up, and their total length is intuitively equal to the difference between the model output value and the baseline value. f ( x )− E [ f This clearly demonstrates the contribution and cumulative effect of each dimension to the final risk value, forming a standard waterfall chart visualization.

[0087] Figure 4 This is a schematic diagram of a waterfall diagram in an embodiment of this application. See also... Figure 4 As shown, the vertical axis lists the feature dimensions sorted by contribution (e.g., "complaint rate," "number of auditors") from bottom to top. The zero point of the horizontal axis represents the baseline expected value of the model output, and each feature corresponds to a horizontal bar. A bar extending to the right indicates that the Shapley value of that feature is positive. ϕi A value >0 indicates a higher risk, and the length of the bar represents the contribution value. A bar extending to the left indicates a negative Shapley value. ϕi A value less than 0 indicates a lower risk. The cumulative total length of all bars represents the current risk prediction. f ( x ) and average benchmark value E [ f The total deviation allows regulators to readily identify key risk drivers and their extent of impact.

[0088] S352: Output waterfall chart.

[0089] After the waterfall chart is generated, it is reviewed by regulatory personnel (the supervisory body). When making selections based on the waterfall chart, regulatory personnel first observe the length and direction of the bars to identify key risk dimensions that extend longer to the right (i.e., have a larger positive contribution). Then, combining their own regulatory experience and business priorities, they manually add or remove target dimensions through interactive operations (such as clicking to check the bars in the chart) based on the automatically selected positive contribution dimensions. For example, if the automatic chart lists the first 5 dimensions, but the regulatory personnel believe that one dimension (such as the complaint rate) needs to be prioritized in the current regulatory cycle, while another dimension (such as the certificate growth rate) can be postponed due to industry context, they can adjust their selection accordingly. Ultimately, they confirm the set of dimensions they have selected as the target dimensions for this rectification, achieving precise decision-making through human-machine collaboration.

[0090] S353: Receive multiple dimensions selected based on the waterfall chart and determine the selected multiple dimensions as the target dimensions.

[0091] The choices here come from proactive and explicit interactive instructions from users (such as regulators), rather than from automatic inference.

[0092] After the waterfall chart is output, the system enters an interactive waiting state until the user submits a specific set of dimensions through a specified action such as clicking, selecting, or confirming the list. The system then directly and unconditionally designates this set of submitted dimensions, regardless of their number or whether they perfectly match the automatically recommended list, as the target dimensions and passes them on to subsequent optimization. This ensures that the final decision on the target dimensions rests with the user, avoiding potential deviations or unclear attribution of responsibility that might arise from automated processing.

[0093] S36: Based on the feature vectors, risk prediction models, constraints, and cost weights corresponding to the target dimension, and with the goal of minimizing the total cost of the optimal modification, counterfactual deduction is performed to obtain the optimal dimension in the target dimension and its corresponding perturbation vector.

[0094] The cost weight is used to indicate the cost required to make changes to each dimension.

[0095] Cost weight W It is a diagonal matrix whose diagonal elements Wii Quantifying the changes i The relative cost or difficulty required for each feature is designed to align the optimization process with actual management practices. In weighting, indicators involving external approvals or scarce resources are assigned higher weights (e.g., ...). W A certification scope of 10 indicates that expanding the certification scope requires an application to CNCA, which is extremely costly. Internally controllable management measures are assigned a medium weight (e.g., W The number of auditors is 5, indicating that recruiting or allocating full-time auditors incurs certain costs. Routine process optimization is assigned a low weight (e.g., ...). W Document archiving rate = 1 or W An internal training completion rate of 1 indicates that the document archiving and record-keeping are complete, the annual internal training targets are met, and the cost is relatively low, thus achieving the optimization goal of minimizing total cost.

[0096] In specific deduction, the target dimension feature vector x target, risk model f Constraints (immutable constraints and feasible region), cost weights W and safety threshold T Substitute safe into the optimization objective function This is a non-convex, non-smooth constrained optimization problem, typically solved using numerical iterative algorithms (such as projective gradient descent or fast iterative threshold shrinkage). The algorithm starts from the initial perturbation... δ Starting with 0, calculate the gradient or proximal mapping of the objective function in each iteration and update it in the direction of reducing the total cost. δ Simultaneously, projection operations are used to ensure that it always satisfies all business constraints. After iterative convergence, the output optimal perturbation vector is... δ The dimension corresponding to the non-zero component in ∗ is the optimal dimension, and its value indicates the specific amount of change (e.g., ...). δ Personnel * = +2).

[0097] To visualize the abstract optimization process and support manual path selection, the multidimensional search space can be reduced and projected into an optimization path graph containing isocost lines and alternative paths. This allows regulators to intuitively evaluate and select a rectification route, and then calculate the specific disturbance vector in reverse.

[0098] Specifically, step S36 above may include: S361: Based on the feature vectors corresponding to the target dimension, the risk prediction model, the constraints, and the cost weights, an optimization path graph is constructed. Each coordinate axis of the optimization path graph represents a different dimension, and corresponding constraints are set on the corresponding coordinate axes. Points in the optimization path graph indicate the risk values ​​under different feature vectors. The optimization path graph also includes multiple risk contour lines, and multiple paths from the current risk value to the warning threshold are configured on the multiple risk contour lines.

[0099] In the specific construction process, within the low-dimensional subspace spanned by the selected target dimensions (e.g., 2-3 key dimensions), each dimension is treated as a coordinate axis, and its feasible upper and lower bounds are marked on the axis as constraints. Then, dense sampling is performed within this space. For each sampled point (i.e., a specific combination of feature values), the risk prediction model f is called to calculate the corresponding risk value. Based on these discrete points, contour plotting algorithms (such as MarchingCubes or 2D interpolation) are used to generate multiple risk contour lines connecting points with the same risk value. Finally, a path search algorithm (such as Dijkstra's algorithm) is used to find multiple feasible descent paths between the contour lines, from the high-risk point representing the current state to the threshold contour line representing the safe state. These paths and their associated coordinate points are then visualized on a graph.

[0100] For plotting feasible descent paths, a graph search algorithm can be used. Specifically, the risk contour map is discretized into a grid, where each grid node corresponds to a set of feature values ​​and its calculated risk value. Algorithms (such as Dijkstra's algorithm) start with the node representing the current state and start from the node with a risk value lower than T. safe The algorithm searches for paths within a grid, targeting a specific node region. During the search, the weight of an edge is determined by the cost of feature changes between nodes (calculated based on the cost weight W). The algorithm identifies multiple paths from the starting point to the target region with different total costs but continuously decreasing risks. Finally, these paths are connected to the corresponding grid nodes in the graph, visually represented as directional curves or polylines.

[0101] Figure 5 This is a schematic diagram of the optimization path graph in the embodiments of this application. See also: Figure 5As shown in the figure, this diagram visually illustrates the counterfactual search process in the feature space. The diagram uses two key features (such as "number of full-time auditors" and "monthly certificate issuance growth rate") as coordinate axes, forming a two-dimensional decision plane. The current state point is located in the high-risk area, and the target safe area is marked by contour lines representing risk thresholds. The multiple curved paths from the current point towards the safe area represent different feature adjustment strategies (e.g., one path tends towards significant staff increases, while another focuses on strict control of certificate issuance). The length and curvature of these paths metaphorically represent the total management costs and adjustment ranges required for different strategies, vividly illustrating how to explore multiple cost-optimal compliance paths under constraints.

[0102] S362: Output the optimal path graph.

[0103] After the optimal path map is output, regulators perform visual analysis and interactive selection based on it. First, they observe multiple feasible descent paths extending from the current high-risk point to a low-risk safe zone, and comprehensively compare the length (intuitively reflecting the total cost of modification), steepness (reflecting the speed of risk reduction), and coordinate regions traversed by each path (reflecting the specific feature dimensions and magnitudes that need adjustment). Then, through interactive methods such as clicking or selecting boxes, they choose the path that best suits the current regulatory strategy, resource constraints, or the institution's actual situation as the target route.

[0104] S363: Receive the target route selected based on the optimal path graph, and calculate the optimal dimension in the target dimension and its corresponding perturbation vector based on the target route.

[0105] Once regulators select a target route, the system extracts a series of key coordinate points along the route. These coordinates correspond to feature vectors at different rectification stages. The system subtracts the coordinates of the route's endpoint (i.e., the feature value at which the safety threshold is reached) from the coordinates of the starting point representing the current state. The difference is the final total disturbance vector δ∗. The dimensions of the non-zero components in this vector are the optimal dimensions within the target dimensions, and their specific values ​​represent the corresponding disturbance values ​​(such as changes on the auditor count axis).

[0106] For example, suppose in an optimization path graph with "number of auditors" and "monthly certificate issuance growth rate" as coordinate axes, regulators choose a broken line path that first rapidly increases the number of auditors and then gradually decreases the certificate issuance rate. The system extracts the coordinates of the endpoint of this path as (auditors = 25, certificate issuance rate = 5%), and the current state of the starting point as (auditors = 20, certificate issuance rate = 15%). By calculating the difference, the disturbance vector is obtained. δ * = (+5, -10%). Therefore, the optimal dimensions are determined to be "number of auditors" and "monthly certificate issuance growth rate," with the specific perturbation vector being that an increase of 5 auditors would decrease the certificate issuance rate by 10 percentage points.

[0107] Because the optimal perturbation vector is obtained through counterfactual deduction. δ * typically contains multiple non-zero components. Furthermore, mathematically, to reduce risk below a threshold with minimum total cost, optimization algorithms often collaboratively adjust multiple features rather than modifying a single dimension. In real-world business, typical risk mitigation strategies also employ a combination of multi-pronged corrective measures (such as simultaneously increasing staff and optimizing processes) to achieve the most efficient and lowest-cost overall compliance goals.

[0108] To ensure that the generated rectification plan is both time-bound and logically coherent, it is necessary to comprehensively consider time costs, execution dependencies, and urgency, intelligently prioritize and schedule multiple optimal dimensions, and generate time recommendations.

[0109] S37: Determine the time cost of each optimal dimension, the dependencies between each optimal dimension during execution, and the target dimension among multiple optimal dimensions that can achieve emergency risk mitigation; sort the optimal dimensions in order of time cost from low to high; in the sorted optimal dimensions, adjust the dimensions with dependencies to adjacent dimensions from back to front; in the adjusted optimal dimensions, place the target dimension at the front; for the currently sorted optimal dimensions, generate time suggestion information for completion within the positively correlated time period.

[0110] The time cost for each optimal dimension refers to the absolute time expected to be consumed in implementing the corresponding rectification measures. For example, the rectification instruction "add two full-time auditors" involves recruitment, training, and onboarding, and its time cost may be quantified as "90 days." The rectification instruction "reduce the monthly certificate issuance growth rate by 10%" is mainly achieved through internal control, and its time cost may only be "15 days."

[0111] The dependencies between the various optimal dimensions refer to the logical sequential constraints between different rectification measures. For example, "upgrading the certification management system" (dimensional A) is a technical prerequisite for "achieving a 100% archiving rate for electronic records" (dimensional B). Therefore, B depends on A, and A must be completed before B can begin.

[0112] Among multiple optimal dimensions, the target dimension that can achieve emergency risk mitigation refers to those dimensions among multiple corrective measures that can quickly and significantly reduce the current core risk. For example, among the two risks of "high complaint rate" and "lack of internal training," the former is directly related to regulatory penalties and reputational damage. Therefore, corrective measures to "reduce the complaint rate" are usually marked by the system or manually as an urgent target dimension that needs to be prioritized.

[0113] When sorting the optimal dimensions, they are first arranged in ascending order according to their estimated time cost, forming an initial sequence. Next, this sequence is scanned from back to front. If a dimension is found to be a predecessor of a dimension earlier in the sequence, that dependent dimension is moved to the position adjacent to the dependent dimension. Finally, regardless of time cost, all dimensions marked as urgent mitigation targets are moved to the front of the sequence. Based on this final sequence, time-based recommendations are generated for each dimension based on its time cost.

[0114] For example, suppose there are three optimal dimensions: A (improve documentation, 5 days, non-urgent), B (recruit auditors, 90 days, dependent on C), and C (reduce certification rate in high-risk areas, 15 days, urgent). First, sort by time cost as [A(5 days), C(15 days), B(90 days)]. A scan reveals that B is a dependency of C, so B is moved before C, resulting in [A(5 days), B(90 days), C(15 days)]. Since C is the urgent objective dimension, it is placed first, resulting in the final sorting as [C(urgent, 15 days), A(5 days), B(90 days)]. Based on this, the following recommendation is generated: "Immediately initiate (Week 1): Reduce certification rate in high-risk areas; Follow-up (1-3 months): Improve documentation archiving and simultaneously initiate the auditor recruitment process." This sorting process generates a highly operable phased rectification roadmap that considers both execution efficiency and logical order, prioritizes core risks, and addresses them first.

[0115] S38: Convert the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, and generate a risk rectification report including risk value, dimension with positive Shapley value, natural language and time suggestion information.

[0116] The final risk rectification report is shown below: Report Number: R-2025-12-001 Certification Entity: XX Quality Certification Center Assessment Date: December 1, 2025 I. Summary of Risk Status Current risk level: 0.68 (High risk) Safety risk threshold: 0.30 Risk level: High risk II. Key Risk Attribution Analysis (Based on Shapley Value) 1. Excessively high monthly certificate issuance growth rate: Contribution value +0.25 2. Insufficient number of full-time auditors: Contribution value +0.18 3. Low customer complaint handling timeliness rate: Contribution value +0.12 III. Recommendations for the Most Cost-Effective Rectification Path 1. Emergency measures (recommended to be implemented immediately, with results expected within 1-2 weeks) Measures: Reduce the monthly certificate issuance growth rate by 15 percentage points.

[0117] Expected effect: This measure can contribute approximately 0.20 to the risk value reduction, and is key to quickly mitigating risk.

[0118] Timeline suggestion: Please adjust your business schedule and complete the deployment by the end of this month.

[0119] 2. Core measures (It is recommended to start them simultaneously and complete them within 1-3 months) Measures: Add two full-time auditors with relevant qualifications.

[0120] Dependency Note: This item is fundamental to ensuring the quality of subsequent audits and compliance expansion. It is recommended to initiate the recruitment process first.

[0121] Timeline suggestion: It is recommended to complete staff onboarding and training before the next quarter.

[0122] 3. Supporting measures (continuous improvement is recommended) Measures: Optimize internal processes to shorten the average processing time for customer complaints to within 5 business days.

[0123] Timeframe recommendation: It is recommended to revise the process within one month and continue monitoring.

[0124] IV. Expected Rectification Results If the above recommendations are fully implemented, the projected risk value will drop to 0.28, below the safety threshold.

[0125] This concludes the explanation of the interpretable attribution and optimization-based certification risk monitoring decision-making method provided in the embodiments of this application.

[0126] Based on the same inventive concept, embodiments of this application also provide an interpretable attribution and optimization certification risk monitoring and decision-making device.

[0127] Figure 6 This is a schematic diagram of the structure of the certification risk monitoring and decision-making device that can explain attribution and optimization in the embodiments of this application. Figure 1 See Figure 6 As shown, the device may include: The vector acquisition module 61 is used to acquire multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification subject.

[0128] The risk prediction module 62 is used to obtain the risk value corresponding to the multi-dimensional feature vector through the risk prediction model. The risk prediction model is a pre-trained machine learning model that can generate the risk value of the certification subject based on the basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information of the certification subject.

[0129] The attribution explanation module 63 is used to generate the Shapley value corresponding to each dimension in the multidimensional feature vector based on the feature vector and the risk prediction model if the risk value is greater than the warning threshold, and select one or more target dimensions to be adjusted based on the Shapley value.

[0130] The optimal deduction module 64 is used to perform counterfactual deduction based on the feature vectors, risk prediction models and constraints corresponding to the target dimension, with the goal of optimal modification, to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The constraints are that the feature vectors of the dimensions of immutable information remain unchanged and the perturbation vectors conform to the business logic.

[0131] The decision report module 65 is used to convert the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, and generate a risk rectification report including risk value, dimension with positive Shapley value, and natural language.

[0132] Furthermore, as a response to Figure 6 In addition to the refinement and expansion of the illustrated device, this application embodiment also provides an interpretable attribution and optimization certification risk monitoring and decision-making device.

[0133] Figure 7 This is a schematic diagram of the structure of the certification risk monitoring and decision-making device that can explain attribution and optimization in the embodiments of this application. Figure 2 See Figure 7 As shown, the device may include: The vector acquisition module 71 is used to acquire multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification subject.

[0134] The future prediction module 72 is used to predict the future feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information of the certification subject after a preset time interval, based on multi-dimensional feature vectors.

[0135] The risk prediction module 73 is used to obtain the risk value corresponding to the future feature vector through the risk prediction model.

[0136] The attribution explanation module 74 is used to generate the Shapley value corresponding to each dimension in the multi-dimensional feature vector based on the feature vector and risk prediction model if the risk value is greater than the warning threshold. The module sorts the multiple dimensions in descending order of Shapley value and selects a preset number of dimensions with positive Shapley values ​​at the top of the sort as the target dimensions.

[0137] The attribution explanation module 74 is specifically used to construct a waterfall plot based on the Shapley value corresponding to the feature vector of each dimension according to the sorting of multiple dimensions. The horizontal axis of the waterfall plot represents the Shapley value, and the vertical axis represents each dimension. It outputs the waterfall plot and receives multiple dimensions selected based on the waterfall plot, and determines the selected multiple dimensions as the target dimensions.

[0138] The optimal deduction module 75 is used to perform counterfactual deduction based on the feature vectors, risk prediction models, constraints and cost weights corresponding to the target dimension, with the goal of minimizing the total cost of the optimal modification, to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The cost weight is used to indicate the cost required to modify each dimension.

[0139] The optimal deduction module 75 is specifically used to construct an optimal path graph based on the feature vectors corresponding to the target dimension, the risk prediction model, constraints, and cost weights. Each coordinate axis of the optimal path graph represents a different dimension, and corresponding constraints are set on the corresponding coordinate axes. Points in the optimal path graph indicate the risk values ​​under different feature vectors. The optimal path graph also includes multiple risk contour lines, and multiple routes from the current risk value to the warning threshold are configured on the multiple risk contour lines. The module outputs the optimal path graph; receives the target route selected based on the optimal path graph; and calculates the optimal dimension in the target dimension and its corresponding perturbation vector based on the target route.

[0140] When there are multiple optimal dimensions, the time suggestion module 76 is used to determine the time cost of each optimal dimension, the dependencies between each optimal dimension during execution, and the target dimension among the multiple optimal dimensions that can achieve emergency risk mitigation; sort the optimal dimensions in order of time cost from low to high; in the sorted optimal dimensions, adjust the dimensions with dependencies to be adjacent from back to front; in the adjusted optimal dimensions, place the target dimension at the front; and generate time suggestion information for completion within the positively correlated time period for the currently sorted optimal dimensions.

[0141] The decision report module 77 is used to convert the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, and generate a risk rectification report including risk value, dimension with positive Shapley value, natural language and time suggestion information.

[0142] It should be noted that the description of the above device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0143] Based on the same inventive concept, this application also provides a computer device.

[0144] Figure 8 This is a schematic diagram of the structure of the computer device in an embodiment of this application. See also... Figure 8 As shown, the computer device may include: a memory 81, a processor 82, and a computer program stored on the memory 81, wherein the processor 82 executes the computer program to implement the methods described in the foregoing embodiments.

[0145] It should be noted that the description of the above computer device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects. For technical details not disclosed in the computer device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0146] Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the methods described in the foregoing embodiments.

[0147] It should be noted that the description of the above computer-readable storage medium embodiments is similar to the description of the above method embodiments, and has similar beneficial effects. For technical details not disclosed in the computer-readable storage medium embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0148] Based on the same inventive concept, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the methods described in the foregoing embodiments.

[0149] It should be noted that the descriptions of the above computer program product embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the computer program product embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0150] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A certification risk monitoring decision-making method that allows for interpretable attribution and optimization, characterized in that, The method includes: Obtain multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification entity; The risk value corresponding to the multidimensional feature vector is obtained through a risk prediction model, wherein the risk prediction model is a pre-trained machine learning model that can generate the risk value of the certification subject based on the certification subject's basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information. If the risk value is greater than the warning threshold, the Shapley value formula is used to generate the Shapley value corresponding to each dimension in the multidimensional feature vector based on the feature vector and the risk prediction model, and one or more target dimensions to be adjusted are selected based on the Shapley value. Based on the feature vector corresponding to the target dimension, the risk prediction model, and the constraints, counterfactual deduction is performed with the goal of optimal modification to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The constraints are that the feature vector of the dimension with immutable information remains unchanged and the perturbation vector conforms to the business logic. The optimal dimension and its corresponding perturbation vector are converted into natural language to indicate the direction of rectification, generating a risk rectification report that includes the risk value, the dimension with a positive Shapley value, and the natural language.

2. The method according to claim 1, characterized in that, The step of performing counterfactual inference based on the feature vector corresponding to the target dimension, the risk prediction model, and the constraints, with the goal of optimal modification, to obtain the optimal dimension in the target dimension and its corresponding perturbation vector includes: Based on the feature vector corresponding to the target dimension, the risk prediction model, the constraints and cost weights, counterfactual deduction is performed with the goal of minimizing the total cost of the optimal modification, to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The cost weights are used to indicate the cost required for each dimension in the modification.

3. The method according to claim 2, characterized in that, Based on the feature vector corresponding to the target dimension, the risk prediction model, the constraints, and the cost weights, and with the goal of minimizing the total cost of the optimal modification, counterfactual deduction is performed to obtain the optimal dimension and its corresponding perturbation vector within the target dimension, including: Based on the feature vector corresponding to the target dimension, the risk prediction model, the constraints, and the cost weights, an optimization path graph is constructed. Each coordinate axis of the optimization path graph represents a different dimension, and corresponding constraints are set on the corresponding coordinate axes. Points in the optimization path graph indicate risk values ​​under different feature vectors. The optimization path graph also includes multiple risk contour lines, and multiple lines from the current risk value to the warning threshold are configured on the multiple risk contour lines. Output the optimization path graph; Receive the target route selected based on the optimization path map, and calculate the optimal dimension and its corresponding perturbation vector in the target dimension based on the target route.

4. The method according to any one of claims 1 to 3, characterized in that, The selection of one or more target dimensions to be adjusted based on Shapley values ​​includes: Sort the multiple dimensions in descending order of Shapley values; Select the first-ranked dimension with a preset number of positive Shapley values ​​as the target dimension.

5. The method according to claim 4, characterized in that, The selection of a predetermined number of dimensions with positive Shapley values ​​at the top of the sorted list as the target dimension includes: A waterfall plot is constructed based on the Shapley values ​​corresponding to the feature vectors of each dimension, according to the sorting of multiple dimensions. The horizontal axis of the waterfall plot represents the Shapley value, and the vertical axis represents each dimension. Output the waterfall diagram; Receive multiple dimensions selected based on the waterfall chart, and determine the selected multiple dimensions as the target dimensions.

6. The method according to any one of claims 1 to 3, characterized in that, Before obtaining the risk value corresponding to the multidimensional feature vector through the risk prediction model, the method further includes: Based on the multidimensional feature vector, predict the future feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification subject after a preset time interval; The step of obtaining the risk value corresponding to the multidimensional feature vector through the risk prediction model includes: The risk value corresponding to the future feature vector is obtained through a risk prediction model.

7. The method according to any one of claims 1 to 3, characterized in that, The number of optimal dimensions is multiple; before generating a risk remediation report including the risk value, dimensions with positive Shapley values, and the natural language, the method further includes: Determine the time cost of each optimal dimension, the dependencies between each optimal dimension during execution, and the target dimension among multiple optimal dimensions that can achieve emergency risk mitigation; Sort the optimal dimensions in order of increasing time cost; In the optimal dimensions after sorting, adjust the dimensions with dependencies to be adjacent from back to front; In the adjusted optimal dimensions, the target dimension is placed first; For the optimal dimension of the current ranking, generate time suggestion information for completion within the positively correlated time period; The generated risk remediation report includes the risk value, dimensions with positive Shapley values, and the natural language, and includes: Generate a risk remediation report that includes the risk value, dimensions with positive Shapley values, the natural language, and the time-based suggestion information.

8. A certification risk monitoring and decision-making device with interpretable attribution and optimization capabilities, characterized in that, The device includes: The vector acquisition module is used to acquire multi-dimensional feature vectors corresponding to the basic information, qualification information, certificate issuance information, personnel information, supervision and feedback information, certification activity information, and innovation activity information of the certification entity; The risk prediction module is used to obtain the risk value corresponding to the multidimensional feature vector through the risk prediction model. The risk prediction model is a pre-trained machine learning model that can generate the risk value of the certification subject based on the certification subject's basic information, qualification information, certificate issuance information, personnel information, supervision feedback information, certification activity information, and innovation activity information. The attribution explanation module is used to generate a Shapley value for each dimension of the multidimensional feature vector based on the feature vector and the risk prediction model if the risk value is greater than the warning threshold, and select one or more target dimensions to be adjusted based on the Shapley value. The optimal deduction module is used to perform counterfactual deduction with the goal of optimal modification based on the feature vector corresponding to the target dimension, the risk prediction model, and the constraints, to obtain the optimal dimension in the target dimension and its corresponding perturbation vector. The constraints are that the feature vector of the dimension with immutable information remains unchanged and the perturbation vector conforms to the business logic. The decision report module is used to convert the optimal dimension and its corresponding perturbation vector into natural language to indicate the direction of rectification, and generate a risk rectification report including the risk value, dimensions with positive Shapley values, and the natural language.

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

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images