CN-122022449-A - Authentication risk supervision decision-making method and device capable of explaining attribution and optimization

CN122022449ACN 122022449 ACN122022449 ACN 122022449ACN-122022449-A

Abstract

The application provides an authentication risk supervision decision method and device capable of explaining attribution and optimization, the method comprises the steps of obtaining basic information, qualification information, verification certificate information, personnel information, supervision feedback information, authentication activity information and multidimensional feature vectors corresponding to innovation activity information of an authentication main body, obtaining risk values corresponding to the multidimensional feature vectors through a risk prediction model, generating Shapley values corresponding to each dimension in the multidimensional feature vectors by adopting a Shapley value formula if the risk values are larger than an early warning threshold value based on the feature vectors and the risk prediction model, carrying out counterfactual deduction with optimal modification as a target based on the feature vectors corresponding to a target dimension, the risk prediction model and constraint conditions, converting the optimal dimension and disturbance vectors corresponding to the optimal dimension into natural language for indicating a rectifying direction, and generating a risk rectifying report. An interpretable and operable supervision closed loop from risk early warning to accurate management is realized.

Inventors

MA YUFEI
TIAN WENTAO
YUAN RUIFENG
Zheng Tuowei
ZHANG JIAN

Assignees

中国网络安全审查认证和市场监管大数据中心

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. A method of authentication risk supervision decision making for interpretive attribution and optimization, the method comprising: acquiring basic information, qualification information, certification issuing information, personnel information, supervision feedback information, authentication activity information and multidimensional feature vectors corresponding to innovation activity information of an authentication main body; Obtaining a risk value corresponding to the multidimensional feature vector through a risk prediction model, wherein the risk prediction model is a machine learning model which is trained in advance and can generate the risk value of the authentication subject based on basic information, qualification information, verification certificate information, personnel information, supervision feedback information, authentication activity information and innovation activity information of the authentication subject; If the risk value is greater than the early warning threshold value, a Shapley value formula is adopted, a Shapley value corresponding to each dimension in the multidimensional feature vector is generated 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; Performing inverse fact deduction by taking optimal modification as a target based on the feature vector corresponding to the target dimension, the risk prediction model and constraint conditions to obtain the optimal dimension and a disturbance vector corresponding to the optimal dimension in the target dimension, wherein the constraint conditions are that the dimension feature vector of the information which cannot be modified is unchanged and the disturbance vector accords with business logic; and converting the optimal dimension and the disturbance vector corresponding to the optimal dimension into natural language for indicating the rectifying direction, and generating a risk rectifying report comprising the risk value, the dimension with the Shapley value being positive number and the natural language.
2. The method according to claim 1, wherein the performing a counterfactual deduction with an optimal modification as a target based on the feature vector corresponding to the target dimension, the risk prediction model and the constraint condition to obtain an optimal dimension and a disturbance vector corresponding to the optimal dimension in the target dimension includes: And carrying out counterfactual deduction by taking the minimum total cost of the optimal change as a target based on the feature vector corresponding to the target dimension, the risk prediction model, the constraint condition and the cost weight, so as to obtain the optimal dimension and the disturbance vector corresponding to the optimal dimension in the target dimension, wherein the cost weight is used for indicating the cost required to be input in the change of each dimension.
3. The method according to claim 2, wherein the obtaining the optimal dimension of the target dimension and the disturbance vector corresponding thereto by performing the counterfactual deduction with the minimum optimal change total cost as a target based on the feature vector corresponding to the target dimension, the risk prediction model, the constraint condition and the cost weight includes: Constructing an optimizing path diagram based on the feature vector corresponding to the target dimension, the risk prediction model, the constraint condition and the cost weight, wherein each coordinate axis of the optimizing path diagram represents different dimensions, the corresponding constraint condition is arranged on the corresponding coordinate axis, points in the optimizing path diagram indicate risk values under different feature vectors, the optimizing path diagram further comprises a plurality of risk contour lines, and a plurality of lines from the current risk value to an early warning threshold value are arranged on the plurality of risk contour lines; Outputting the optimizing path diagram; And receiving a target line selected based on the optimizing path diagram, and calculating to obtain the optimal dimension and the disturbance vector corresponding to the optimal dimension based on the target line.
4. A method according to any one of claims 1 to 3, wherein the selecting one or more target dimensions to be adjusted based on Shapley values comprises: sorting the plurality of dimensions according to the sequence from large to small of the Shapley value; And selecting a preset number of dimensions with the forefront ranking and Shapley value as positive numbers as the target dimensions.
5. The method of claim 4, wherein selecting a preset number of forward-ordered Shapley positive dimensions as the target dimension comprises: Constructing a waterfall diagram based on the shape values corresponding to the feature vectors of each dimension according to the ordering of the plurality of dimensions, wherein the abscissa of the waterfall diagram represents the shape values and the ordinate represents the dimensions; Outputting the waterfall diagram; And receiving a plurality of dimensions selected based on the waterfall graph, and determining the selected plurality of dimensions as the target dimension.
6. A method according to any one of claims 1 to 3, wherein before obtaining risk values for the multi-dimensional feature vectors by a risk prediction model, the method further comprises: Based on the multidimensional feature vector, predicting future feature vectors corresponding to basic information, qualification information, certification verification information, personnel information, supervision feedback information, authentication activity information and innovation activity information of the authentication subject after a preset time interval; the obtaining, by the risk prediction model, a risk value corresponding to the multidimensional feature vector includes: and obtaining a risk value corresponding to the future feature vector through a risk prediction model.
7. The method of any one of claims 1 to 3, wherein the number of optimal dimensions is a plurality, and wherein prior to generating a risk modification report comprising the risk value, a dimension having a Shapley value that is a positive number, and the natural language, the method further comprises: determining the time cost of each optimal dimension, the dependency relationship of each optimal dimension when the optimal dimension is executed, and the target dimension which can realize the risk emergency relief in the multiple optimal dimensions; sequencing all the optimal dimensions according to the sequence from low to high of the time cost; In the ordered optimal dimension, the dimension with the dependency relationship is adjusted to be adjacent from back to front; In the adjusted optimal dimension, the target dimension is placed at the forefront; Generating time suggestion information completed in a positive correlation time period aiming at the optimal dimension of the current sequencing; The generating a risk modification report including the risk value, the dimension with the Shapley value being a positive number, and the natural language includes: and generating a risk modification report comprising the risk value, the dimension of which the Shapley value is positive, the natural language and the time suggestion information.
8. An authentication risk supervision decision making apparatus interpretable attribution and optimization, the apparatus comprising: the vector acquisition module is used for acquiring multidimensional feature vectors corresponding to basic information, qualification information, certification verification information, personnel information, supervision feedback information, authentication activity information and innovation activity information of the authentication main body; The risk prediction module is used for obtaining a risk value corresponding to the multidimensional feature vector through a risk prediction model, wherein the risk prediction model is a machine learning model which is trained in advance and can generate the risk value of the authentication subject based on basic information, qualification information, certification verification information, personnel information, supervision feedback information, authentication activity information and innovation activity information of the authentication subject; the attribution interpretation module is used for generating a shape value corresponding to each dimension in the multi-dimensional feature vector by adopting a shape value formula based on the feature vector and the risk prediction model if the risk value is larger than an early warning threshold value, and selecting one or more target dimensions to be adjusted based on the shape value; The deduction optimal module is used for carrying out counterfactual deduction by taking optimal change as a target based on the feature vector corresponding to the target dimension, the risk prediction model and the constraint condition to obtain the optimal dimension and the disturbance vector corresponding to the optimal dimension in the target dimension, wherein the constraint condition is that the dimension feature vector of the non-changeable information is unchanged and the disturbance vector accords with service logic; And the decision report module is used for converting the optimal dimension and the disturbance vector corresponding to the optimal dimension into natural language for indicating the rectifying direction and generating a risk rectifying report comprising the risk value, the dimension with the shape value being a positive number and the natural language.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, 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 the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Description

Authentication risk supervision decision-making method and device capable of explaining attribution and optimization Technical Field The present application relates to the field of data processing technology, and in particular, to an authentication risk supervision decision method capable of interpreting attribution and optimizing, an authentication risk supervision decision apparatus capable of interpreting attribution and optimizing, a computer device, a computer readable storage medium, and a computer program product. Background The supervision and management body refers to an organization, such as an administrative supervision organization, an administrative self-administration organization, etc., which performs supervision, inspection, evaluation, guidance and if necessary, sanctions on compliance, quality, safety or behavior of a specific object (such as an enterprise, an organization, a person, a product, an activity, etc.) under a management system. The certification body refers to a professional organization or organization, such as each quality certification center, that evaluates the compliance of products, services, management systems, personnel, etc., and issues certification certificates to prove compliance thereof, according to specific standards, specifications, and technical procedures. In order to prevent the certification body from damaging the certification credit due to insufficient capacity, illegal operation or credit loss, thereby damaging the rights and interests of consumers and disturbing the industrial order, a supervision and management body is required to conduct risk supervision on the certification body. At present, a supervision and management body performs risk supervision on an authentication body, and a risk prediction model (such as a random forest, a gradient lifting decision tree or a deep neural network and other algorithms) based on machine learning is mainly adopted. Specifically, first, historical data of a plurality of authentication subjects are collected, risk value marks are carried out on each authentication subject, and then the historical data of the plurality of authentication subjects and risk values thereof are used as training sets. Then, the machine learning model is trained using the training set. Finally, aiming at the authentication subject to be supervised, acquiring relevant data of the authentication subject to be supervised, and inputting the acquired relevant data into a trained machine learning model. The model outputs a specific value, and the data is the risk value of the authentication subject to be supervised. The supervision and management body realizes supervision on the authentication body based on the risk value. However, the machine learning model is used as a black box, only provides a risk value, cannot reveal specific causes and responsibilities of risks, causes lack of interpretation and convincing of supervision decisions, and faces the supervision problem of being based on opacity. Moreover, the machine learning model only stops at risk early warning, and can not generate an rectification guide with operability based on attribution results, so that a supervision and management body is difficult to guide an authentication body to effectively reduce risks. Therefore, there is a need to provide an interpretable and operable regulatory decision method to penetrate a risk black box to achieve paradigm upgrades from pre-warning to abatement. Disclosure of Invention It is an object of embodiments of the present application to provide an interpretable attribution and optimized authentication risk supervision decision method, an interpretable attribution and optimized authentication risk supervision decision apparatus, a computer device, a computer readable storage medium, and a computer program product, to provide an interpretable, operable supervision decision report, to implement a paradigm upgrade of authentication supervision from pre-warning to governance. In order to solve the technical problems, the embodiment of the application provides the following technical scheme: The first aspect of the application provides an authentication risk supervision decision method capable of interpreting attribution and optimization, which comprises the steps of obtaining basic information, qualification information, issuing certification information, personnel information, supervision feedback information, authentication activity information and multidimensional feature vectors corresponding to innovation activity information of an authentication main body, obtaining risk values corresponding to the multidimensional feature vectors through a risk prediction model, wherein the risk prediction model is a machine learning model which is trained in advance and can generate the risk values of the authentication main body based on the basic information, the qualification information, the issuing certification information, the personnel information, the