Decision intervention explanation optimization method and system based on human and social big model decision

Abstract

The invention discloses a decision intervention explanation optimization method and system based on human-social big model decision, and relates to the technical field of information systems, and the method comprises the steps: extracting multi-dimensional heterogeneous features based on multi-source data of a human-social business system, carrying out the preliminary weight distribution of the features of each dimension, screening key data, and carrying out the optimization of the decision intervention explanation. Inputting the initial weight into an adjustment model to dynamically adjust the initial weight to obtain a scenarized explanation weight matrix; obtaining an original decision trajectory of the large model reasoning path and the robot process automatic execution trajectory, and inputting the original decision trajectory and the scenarized explanation weight matrix into an intervention generation adversarial evolution network to generate an intervention sample; key nodes are identified from the original decision trajectory of the robot, and the basic interpretation vectors of the key nodes are weighted and summed according to the comprehensive weight to obtain enhanced interpretation vectors; a decision report is generated based on the enhanced interpretation vector and the intervention sample. The problem of interpretability of a large model decision'black box 'and an RPA mechanical execution track is solved, and the compliance and accuracy of a decision system are improved through dynamic optimization.

Description

Technical Field

[0001] This invention relates to the field of information systems technology, and in particular to a method and system for decision intervention, interpretation and optimization based on a large human resources and social security model. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] With the accelerated digital transformation of human resources and social security services, intelligent decision-making based on large models and mechanical execution by RPA (Robotic Process Automation) have become key supports for core business processes. However, the "black box" nature of intelligent decision-making means that its internal logic and decision-making basis lack transparency, leading to phenomena such as skipping necessary verification nodes. This makes it difficult to rely on these models, which lack clear explanatory capabilities, for decision review in actual business scenarios when there are discrepancies between automated approval results and actual policy implementation.

[0004] Because human resources and social security operations require not only efficient and precise automated processing, the "black box" nature of large-scale models and the unexplainable nature of RPA trajectories make it difficult to trace the decision-making process, locate abnormal operations, and verify policy compliance. This severely restricts the credibility of business systems and regulatory compliance requirements, necessitating the introduction of manual review mechanisms at key stages to ensure the accuracy of decisions and legal compliance. Therefore, if large-scale models cannot provide sufficient explanatory information, human reviewers will be unable to fully understand the rationale behind decisions, increasing the difficulty of auditing and error correction, and making it difficult to clarify responsibility and corrective measures in the event of disputes.

[0005] Moreover, as an automated process execution tool, current RPA can quickly and in batches output operations based on these models. However, if the lack of transparency in the decision-making logic leads to errors, the impact will be amplified rapidly, further increasing the workload of manual review. It may also lead to slow response times in the approval process, difficulty in tracing and correcting errors, and pose serious risks to regulatory audits, public trust, and legal compliance. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this invention provides a method and system for decision intervention interpretation and optimization based on a large human resources and social security model. It aims to solve the interpretability problems of the "black box" of large model decision-making and the mechanical execution trajectory of RPA, and improve the compliance and accuracy of the decision-making system through dynamic optimization.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, this invention provides a decision intervention interpretation and optimization method based on a large-scale human resources and social security model, including: Acquire multi-source data from the human resources and social security business system, extract multi-dimensional heterogeneous features based on the multi-source data, vectorize the multi-dimensional heterogeneous features, and form a multi-dimensional interpretation and representation space; Preliminary weight allocation is performed on the features of each dimension in the multidimensional interpretation representation space. Key data is screened and input into the adjustment model to dynamically adjust the preliminary weights, resulting in an updated scenario-based interpretation weight matrix. The original decision trajectories of the large model reasoning path and the automated execution trajectory of the robot process are obtained and input into the intervention generative adversarial evolution network along with the scenario-based interpretation weight matrix to generate intervention samples. Key nodes are identified from the original decision trajectories of the robotic process automation execution trajectory. Based on the adversarial contribution of the key nodes and the scenario-based interpretation weight matrix, a comprehensive weight is obtained. The basic interpretation vectors of the key nodes are weighted and summed according to the comprehensive weight to obtain the enhanced interpretation vector. Based on the enhanced explanatory vector and intervention samples, a decision report is generated, which presents the decision intervention logic, decision process and optimization results from multiple dimensions.

[0008] Further technical solutions include multi-dimensional heterogeneous features such as policy provisions, risk factors, business logic rules, historical case matching degree, and real-time data consistency.

[0009] Further technical solutions include a database of historical misjudgment cases, policy change logs, and real-time risk status.

[0010] A further technical solution is that the intervention-generated adversarial evolutionary network includes a generator, a discriminator, and an adversarial perturbation module. The adversarial perturbation module guides the generator to prioritize perturbing high-weight features based on a scenario-based interpretation of the weight matrix.

[0011] A further technical solution is that the generation of intervention samples specifically involves: A well-trained generator and discriminator are obtained through two-stage training, and preliminary intervention samples are generated based on the well-trained generator. The initial intervention sample was optimized through evolutionary tuning using policy rules and constraints to obtain the optimized intervention sample.

[0012] Further technical solutions include policy and rule constraints such as mandatory hard constraints, guiding soft constraints, conditional constraints, and priority constraints.

[0013] A further technical solution involves evaluating the optimized intervention samples based on evaluation metrics to obtain evaluation results; retraining the intervention generative adversarial evolution network based on the evaluation results to obtain an optimized intervention generative adversarial evolution network; obtaining optimized intervention samples based on the optimized intervention generative adversarial evolution network; and generating a decision report based on the reinforcement explanatory vector and the optimized intervention samples.

[0014] Secondly, the present invention provides a decision intervention interpretation and optimization system based on a large-scale human resources and social security model, characterized in that it includes: The interpretation representation construction module is configured to: acquire multi-source data from the human resources and social security business system, extract multi-dimensional heterogeneous features based on the multi-source data, vectorize the multi-dimensional heterogeneous features, and form a multi-dimensional interpretation representation space. The adaptive weight adjustment module is configured to: perform preliminary weight allocation on the features of each dimension in the multidimensional interpretation representation space, filter key data, input the key data into the adjustment model to dynamically adjust the preliminary weights, and obtain the updated scenario-based interpretation weight matrix. The intervention sample generation module is configured to: obtain the original decision trajectory of the large model reasoning path and the robot process automated execution trajectory, and input it together with the scenario interpretation weight matrix into the intervention generation adversarial evolution network for processing to generate intervention samples; The interpretation enhancement module is configured to: identify key nodes from the original decision trajectory of the robot process automation execution trajectory, obtain a comprehensive weight based on the adversarial contribution of the key nodes and the scenario-based interpretation weight matrix, and sum the basic interpretation vectors of the key nodes according to the comprehensive weight to obtain the enhanced interpretation vector; The decision report generation module is configured to generate a decision report based on the enhanced explanatory vector and intervention samples, and to display the decision intervention logic, decision process and optimization results from multiple dimensions.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the decision intervention interpretation and optimization method based on the human resources and social security big model as described in the first aspect.

[0016] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the decision intervention interpretation and optimization method based on the human resources and social security big model as described in the first aspect.

[0017] The above one or more technical solutions have the following beneficial effects: This invention addresses the "black box" nature of large models and the uninterpretable nature of RPA trajectories by proactively capturing and comprehensively understanding key review nodes in the automated decision-making process. It constructs a clear thought process chain, enabling human reviewers to quickly get started and intervene efficiently. Specifically, this invention constructs a multi-dimensional interpretive representation space, designs scenario-adaptive weight operators, introduces an intervention-generation adversarial evolution mechanism, and combines it with policy and rule constraints to achieve the identification, interpretation, and risk intervention of key nodes in the automated decision-making process. Ultimately, it outputs a chain-like visual interpretive report with high reviewability and traceability.

[0018] This invention achieves multi-dimensional interpretability of the decision-making process, covering core judgment elements such as policy compliance, risk control, and business rule logic; it achieves scenario-adaptive weight adjustment, improving the system's dynamic response capability to policy changes, risk situations, and accumulated misjudgment cases; it achieves the generation of adversarial evolution intervention samples, ensuring the authenticity, validity, and business compliance of the samples; and it achieves the integration of visual reports and manual review, enhancing the system's credibility and regulatory compliance. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of the decision intervention interpretation and optimization method based on the human resources and social security big data model in an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] Example 1 like Figure 1As shown in the figure, this embodiment discloses a decision intervention interpretation and optimization method based on a large human resources and social security model. The method includes the following steps: S1: Obtain multi-source data from the human resources and social security business system, extract multi-dimensional heterogeneous features based on the multi-source data, vectorize the multi-dimensional heterogeneous features, and form a multi-dimensional interpretation and representation space; By collecting data from multiple sources and extracting structured features, the "black box" decision-making logic of large models and the mechanical execution trajectory of RPA are transformed into quantifiable and analyzable structured data, providing raw data support for subsequent steps such as explanatory vector construction and weight calculation.

[0025] S101: Collect full-process trajectory data of the human resources and social security business system to obtain multi-source data; extract multi-dimensional heterogeneous features based on multi-source data.

[0026] We collect end-to-end trajectory data directly related to large-scale model decision-making and RPA execution from the human resources and social security business system, fully reconstructing the "operation chain" of decision-making and execution to ensure sufficient raw information for subsequent feature extraction. Data collection sources mainly include three types of system data: social security administration system, employment service platform, and other related systems.

[0027] (1) Social Security Administration System 1) Record operation logs for core business processes such as pension insurance qualification certification and benefit disbursement, mainly involving data records such as submission time, material review steps, and automatic verification results; 2) Interface call records, mainly involving business service data such as verifying participation status with the medical insurance system and verifying contribution base with the tax system; 3) Data flow path, such as the transmission link of applicant information from the "acceptance module" to the "rule engine module" and then to the "large model decision module".

[0028] (2) Employment service platform 1) Record the entire process data of unemployment insurance claims, employment subsidy disbursements, etc., mainly involving electronic forms submitted by applicants, including declaration information such as reasons for unemployment and duration of social security suspension; 2) Rule verification nodes automatically triggered by the system, such as verifying the feature of "involuntary resignation"; 3) RPA execution trajectory, such as the start and end time of RPA automatically calling the social security database to verify the participation status, and the response results of calling external employment registration interfaces.

[0029] (3) Other related systems The external data required for supplementary decision-making mainly involves: 1) some data from the tax system, used to verify the authenticity of income declarations; 2) some data from the banking system, used to verify payment records; and 3) some data from the audit system, used to store historical cases of insurance fraud, etc.

[0030] Multi-source data collection covers the recording content of both decision-making and execution trajectories, mainly including large model inference logs, RPA execution trajectories, and business system metadata.

[0031] (1) Large Model Inference Log The model receives input data including form fields filled out by the applicant and supporting materials; intermediate reasoning processes, including calculation of the matching degree of policy provisions and output of risk factor scores; and the final decision result, whether to approve or reject, and the confidence level.

[0032] (2) RPA execution trajectory The RPA robot receives task instructions, such as the command "automatically check cross-regional insurance records"; the specific execution steps are from logging into the social security system → entering the insured person's ID → extracting the insurance time → finally returning the result; and the exception handling records include the number of retries after the interface call failed, the manual intervention nodes, and other records.

[0033] (3) Business system metadata The operation steps include timestamps, user identity information, and business type tags, such as "pension insurance qualification certification" and "unemployment insurance benefit application".

[0034] Based on all collected trajectory data, five key features were extracted for core human resources and social security business scenarios: policy basis, risk factors, business logic rules, historical case matching degree, and real-time data consistency. These features cover policy compliance, risk controllability, process logic, historical experience reference, and data reliability, as detailed below: (1) Policy provisions are based on the characteristics of policy compliance. The system uses NLP technology to extract relevant clauses from the social security policy and regulation database. Key extraction fields include clause ID, clause content, applicable condition label, and effective date. For example, in the scenario of "pension insurance qualification certification", the system needs to verify the condition of "accumulated payment for 15 years". The system uses NLP to extract the clause from Article 16 of the "Social Insurance Law of the People's Republic of China" and generates a feature record as [Clause ID 001, 15 years of continuous payment qualify for pension, social security + pension, effective date 2023-01-01].

[0035] (2) Risk factor labeling risk controllability characteristics Based on historical audit data, the potential risks of current cases are quantified, and the risk level is quantified to assist in decision-making regarding whether manual review is required. High-risk cases are automatically marked as "requiring manual verification." Based on historical audit data such as historical insurance fraud case databases and records of irregular supplementary payments, quantifiable risk indicators are extracted, and a risk score (0-1 range, with 1 indicating high risk) is calculated using a logistic regression model. For example, if an applicant has "participated in insurance across regions 3 times in the past 3 months," according to industry experience thresholds, 2 times would be considered abnormal, and the model calculates their risk score to be 0.5.

[0036] (3) Business logic rules mark the logical characteristics of the process. Based on the built-in automated rule nodes and their triggering conditions in the business process, interpret whether the decision conforms to the preset process and verify whether the decision conforms to the business process rules. Parse the configuration file of the human resources and social security business process rule engine, extract the rule nodes and their triggering conditions, and represent the rule node ID, rule description, and triggering condition expression in a structured manner. For example, in the unemployment insurance benefit application process, the rule engine is configured that "unemployment registration requires involuntary resignation + social security cessation of contributions for at least 6 months." Extract the rule node as "Resignation Reason Verification," with the triggering condition: rejection if the resignation reason ≠ 01 / 02 / 03, and "Social Security Cease Payment Duration Verification," with the triggering condition: rejection if the number of days of cessation of contributions is less than 180 days.

[0037] (4) Historical case matching degree markers and historical experience features The similarity between the current case and historical successful / failed cases is calculated to explain whether the decision aligns with historical experience. This leverages historical experience to aid decision-making and improve the efficiency and accuracy of handling new cases. The structured features of the current case are aligned with cases in the historical case database. Cosine similarity is calculated for the features, and a matching score is assigned, linking the decision results and key differences from historical cases. For example, the cosine similarity between the current case "insurance duration 16 years, risk score 0.3" and a case in the historical successful case database "insurance duration 15 years, risk score 0.2" is 0.85, and is labeled "matched successful case, decision should be passed."

[0038] (5) Real-time data consistency markers data reliability characteristics The consistency between the input data of the large model and the authoritative external data source is used to explain whether the "decision data is reliable," identify data input errors or false declarations, and avoid decision-making errors caused by unreliable data. The fields input to the large model are compared with the corresponding fields from the external data source. The proportion of conflicting fields is calculated, and specific conflict points are marked. For example, if an applicant fills in "income of 120,000 yuan in the past 12 months," and the system retrieves the tax system record showing "individual income tax declaration of 100,000 yuan," the conflicting field is "income," and the conflict level is "20%."

[0039] S102: Vectorize multi-dimensional heterogeneous features to form a structured multi-dimensional interpretation representation space.

[0040] The extracted multi-dimensional heterogeneous features are uniformly transformed into low-dimensional, continuous vector representations, forming a structured "multi-dimensional interpretive representation space". This space serves as the input basis for subsequent steps such as weight calculation and sample generation, and also as a "semantic coordinate system" for interpreting the decision logic of large models and the execution trajectory of RPA.

[0041] Each feature dimension is vectorized separately: for policy clauses, a 10-dimensional one-hot vector is mapped using clause ID hashing; the applicable condition expression is encoded into a 20-dimensional vector using syntax tree encoding; and the vectors are concatenated to form a 30-dimensional vector. For risk factors, one-hot encoding of anomaly types is used, risk values ​​are normalized to [0,1], and concatenated into a 6-dimensional vector. To address the characteristics of business logic rules, the rule logic expressions are converted into Reverse Polish Notation, and then encoded into a 15-dimensional vector using a Recurrent Neural Network (RNN). Based on the historical case matching characteristics, the similarity values ​​of the top 3 similar cases are taken, normalized, and concatenated into a 3D vector. To address the consistency characteristics of real-time data, the number of inconsistent fields is normalized to (0-1), the consistency rate is normalized to (0-1), and the results are concatenated into a 2D vector. .

[0042] Concatenate all feature vectors according to their dimensions to form a multidimensional explanatory vector set:

[0043] Each sub-vector corresponds to a feature representation of one dimension.

[0044] The five-dimensional vectors are concatenated into a high-dimensional space, Space, which ultimately outputs a structured multidimensional interpretive representation space:

[0045] Among them, the dimension metadata records the meaning, calculation method and related business rules of each feature.

[0046] S2: Perform preliminary weight allocation on the features of each dimension in the multidimensional interpretation representation space, filter key data, input the key data into the adjustment model to dynamically adjust the preliminary weights, and obtain an updated scenario-based interpretation weight matrix. By quantifying the impact of each dimension of features on the final decision through weight allocation, and continuously optimizing the weights through a dynamic learning mechanism, the decision system can both respond quickly based on prior knowledge and adapt to changes in business scenarios.

[0047] The initial weights are set based on the experience of human resources and social security business experts. For now, the weights for policy provisions are set at 0.4, risk factors at 0.3, business logic rules at 0.2, historical case matching at 0.1, and real-time data consistency at 0.0, forming the initial weight matrix:

[0048] The weights are updated based on real-time decision feedback, enabling the decision-making system to adapt to revisions in business policies and changes in rules. The feedback data mainly includes the behavior logs of downstream systems such as manual review and correction, new case results, manual review results, user complaint rate, and subsequent manual review time.

[0049] The core of dynamic adjustment is based on the weight adjustment amount. The weight increment (or decrement) that changes over time or in a given scenario represents the weight change over time. Compared to The change in the weights of each dimension at that time.

[0050] It is a model learning driven by multi-source dynamic input. Based on changes in real-time business scenarios, it mainly relies on three types of key data, namely "historical misjudgment case library", "policy change log", and "real-time risk situation", to dynamically adjust the weights. These three types of key data constitute the input source of scenario perception, and dynamically adjust the weights of each dimension of features in decision interpretation so that the weight allocation is highly matched with the current business needs.

[0051] The above three types of data are input into the adjustment model to trigger weight adjustments. The specific data content and scene awareness functions are as follows: (1) Historical Misjudgment Case Database Record cases of "system decision-making errors" discovered during manual review, including error types (false approval / false rejection) and error reason annotations (ignoring policy clause ID002, risk factor calculation deviation, etc.).

[0052] By analyzing the characteristic distribution of misjudgment cases, we can identify the weak links in the current weight allocation, such as the frequent misjudgments caused by an excessively low weight in a certain dimension.

[0053] Example: If 80% of the incorrectly rejected cases in the past involve the matching degree of historical cases. This indicates that the existing historical case matching weight (0.1) is too low and needs to be increased.

[0054] (2) Policy Change Log Record information on revisions to human resources and social security policies and regulations, including additions, revisions, or repeals of policy clauses; revision type (strongly binding / weakly binding), where a strongly binding clause's effective date requires immediate adjustment of its weight, and a weakly binding guidance opinion can be adjusted gradually; and related dimensions, such as the dimension on which the revision is based. .

[0055] By analyzing policy change logs, we can identify policy-sensitive periods, such as the week before a similar clause takes effect, and trigger weight adjustments related to that policy dimension.

[0056] Example: When the rule for calculating contribution years is detected to have been revised from "continuous" to "cumulative," the policy provisions are based on the following dimensions. The weight needs to be temporarily increased to 0.5 to ensure that the new policy is applied first.

[0057] (3) Real-time risk situation The current risk level is calculated in real time through the risk monitoring system, including the regional risk index. If the number of insurance fraud cases in a certain region increases by 50% month-on-month, it is marked as high risk. Business type risk: If the fraud rate of cross-regional insurance business is higher than the average by 30%, it is marked as high risk business. Time window risk: If the probability of fraud increases during the peak application period at the end of the month, it is marked as a high-risk period.

[0058] By assessing the risk situation in real time, high-risk scenarios can be identified, and weight adjustments can be triggered for dimensions related to risk prevention and control.

[0059] Example: When a region is marked as high-risk, the risk factor dimension The weight needs to be temporarily increased to 0.6 to strengthen the screening of abnormal cases.

[0060] The adjusted model is built using the XGBoost model. Internally, the model automatically processes and fuses features from three types of input data (historical misjudgment case library, policy change logs, and real-time risk situation). It learns the relationship between features and decision trees by constructing multiple decision trees. Relationship: The split nodes of each tree are based on the importance of features (e.g., policy revision type to the relationship). (With a greater impact), the leaf nodes of the tree output the path below. The cumulative value is optimized stepwise through gradient boosting. Finally, the model sums the weighted outputs of all leaf nodes of the trees to obtain the weight adjustments for each dimension. .

[0061] Contextualized interpretation of the weight matrix The weight of the previous time step Combination The core logic of the subsequent update results is incremental adjustment + normalization. The calculation formula is:

[0062] in, Indicates the time step. Normalization is required to ensure that the sum of the weights is 1.

[0063] For every 100 new misclassified cases received, incremental training of the model is triggered, and the model is updated. ,make sure Adapt to the latest business scenarios and special scenarios and periods such as policy revisions.

[0064] S3: Obtain the original decision trajectory of the large model reasoning path and the robot process automated execution trajectory, and input it together with the scenario-based interpretation weight matrix into the intervention generation adversarial evolution network for processing to generate intervention samples; By employing a Generative Adversarial Network (GAN) framework and combining business rules with adversarial learning, this invention generates intervention samples capable of specifically correcting decision-making biases in large models or RPA execution defects. Examples of interventions include correcting certain data fields or inserting necessary verification nodes. This invention innovatively designs an intervention-based generative adversarial evolutionary network (I-EvGAN), which satisfies the requirements of "generated sample authenticity" (close to human correction), "intervention effectiveness" (improving decision-making results), and "business compliance" (compliance with policy provisions). Therefore, it extends the traditional GAN ​​by introducing an "adversarial perturbation module" and "scenario-based weight constraints."

[0065] The core of I-EvGAN consists of three modules: a generator G, a discriminator D, and an adversarial perturbation module. The input is a context-specific interpretation of the weight matrix. The large model inference path and the original decision trajectory of RPA execution trajectory are fed to the generator G, which outputs intervention samples. Real intervention samples, such as manually corrected historical records, are input to the discriminator D, which outputs a authenticity probability (determining whether it is a real sample) and a perturbation score (evaluating the difference between the generated sample and the original sample). The perturbation score is the discriminator D's quantitative assessment of the "difference between the generated sample and the real sample," reflecting the authenticity of the generated sample, i.e., the degree of similarity between the machine and manually corrected samples. The adversarial perturbation module... The feature weights of each dimension guide the generator G to prioritize perturbing the high-weight dimensions, such as... Then the focus should be on adjusting parameters related to risk.

[0066] The raw decision trajectory data of the large model inference path and RPA execution trajectory can be obtained through the model and system built-in log records: The inference path of the large model is its internal processing flow when generating decisions. Some large models support outputting key information in the inference process, which can be obtained directly by configuring model parameters (such as output_hidden_states=True); The execution trajectory of RPA is a time series record of its operation steps. Mainstream RPA tools record execution logs by default, including operation type (click, input, system call), timestamp, interface element coordinates, input data content, etc., which can be directly exported as structured data.

[0067] Generator G: Input the original decision trajectory (Wt, large model inference path, RPA trajectory), output the generated intervention sample (simulating artificially corrected pseudo-samples, i.e., pseudo-intervention samples after machine learning). Discriminator D takes two types of samples as input: real samples (human-modified historical records, decision trajectories actually modified by business experts, such as real intervention data where the original weight Wt=0.3 was modified to 0.5); and generated samples (pseudo-modified data) output by G. D's task is to distinguish between these two types of samples and output the probability of authenticity (determining whether a sample is real) and the perturbation score (evaluating the difference between generated and real samples).

[0068] S301: The specific steps for generating preliminary intervention samples are as follows: Through two-stage training (unsupervised pre-training + adversarial training), a set of real and effective intervention samples covering typical risk scenarios is generated as the "seed input" for subsequent S4, providing supervision signals for the optimization of the generator (G), enabling it to generate more accurate real and effective intervention samples, ultimately serving the goal of correcting decision biases in large models or RPA execution defects.

[0069] (1) First stage (unsupervised pre-training) By utilizing historical manual correction records, the generator G learns the basic pattern of "typical intervention operations" and generates the initial parameters of the pre-trained generator G.

[0070] Input historical manual correction records, which are real intervention cases from the human resources and social security business system. These records mainly include the problem context, correction operation, and correction result. For example, if the problem is "the applicant's social security payment suspension period is only 5 months, which does not meet the 6-month requirement," insert a "Social Security Payment Suspension Period Verification" node, "correct the suspension period to 6 months," and then output the rejection correction result.

[0071] Input the large model inference log and RPA execution trajectory corresponding to the problem context, such as operation records such as input data, rule triggering nodes, and final decision.

[0072] (2) Second stage (competitive training) Training the discriminator D: Input real samples (randomly selected from historical manually corrected records), labeled "real" and generated samples (generated by generator G based on current parameters), labeled "generated", and perform optimization to minimize cross-entropy loss.

[0073] Training the generator G: By optimizing the adversarial loss, we maximize the probability of the discriminator D being true to the generated samples, while minimizing the perturbation score, thereby improving the authenticity of the samples.

[0074] After 100 iterations, when the discriminator D has a probability ≥ 0.8 of the generated samples being genuine, pre-training stops, and a preliminary intervention sample set covering typical risk scenarios such as policy application errors, data inconsistencies, and high-risk operations is output. :

[0075] Each sample is labeled with associated key nodes, such as "policy clause verification node" and "risk scoring node".

[0076] S302: The initial intervention sample is optimized through policy and rule constraints to obtain the optimized intervention sample, specifically: By transforming human resources and social security business rules and policy provisions into formal constraints that can be processed by computers, the system ensures that all intervention operations comply with business norms and policy requirements during the generation and evolution of intervention samples. Constraint coding serves as a bridge connecting "business rules" and "algorithm optimization." Through formally expressing constraints, the system can automatically check compliance when generating, evaluating, and optimizing intervention samples, avoiding the generation of invalid or dangerous samples that violate policies.

[0077] Input data from the human resources and social security policy and regulation database, business operation manual, and historical audit cases (high-frequency violation types), extract key rules, and label rule types and related fields. Rule types are categorized as follows: (1) Mandatory hard constraints Strictly enforced rules will render any intervention sample invalid if violated. For example, policy bottom lines include "applications for unemployment benefits must be rejected if the applicant is under 16 years old"; data legality rules include "the contribution base must be ≥ the local minimum wage standard"; and process compliance rules include "high-risk operations must be subject to manual review."

[0078] (2) Guiding soft constraints The rules suggest compliance but allow exceptions, with penalties (reduced sample weight) for violations. For example, optimization rules such as "when the number of cross-regional insurance enrollments is abnormal, it is recommended to supplement and verify the insurance enrollment records for the past 3 months"; and efficiency rules such as "the decision-making steps for simple business (such as pension qualification certification) should not exceed 5 steps".

[0079] (3) Conditional constraints Rules that only take effect under specific conditions should be considered in conjunction with the context to determine whether they are triggered. For example, rules based on regional differences, such as "If the applicant's place of insurance is City A, then the contribution base must be ≥ the minimum wage of City A"; and rules based on time windows, such as "During the concentrated application period at the end of the month (25th-30th of each month), risk screening needs to be strengthened."

[0080] (4) Priority constraints The importance of different rules is ranked to resolve rule conflicts. For example, in the case of policy conflicts, "when the Social Insurance Law of the People's Republic of China conflicts with local implementation rules, the Social Insurance Law of the People's Republic of China shall prevail"; in the case of risk levels, "compliance constraints for high-risk operations take precedence over efficiency constraints".

[0081] The constraints are encoded, and policies and regulations are parsed using legal text mining tools to extract constraint rules. Business operation manuals are converted into executable rules using a rule engine. Historical audit cases are used to extract high-frequency violation patterns through association rule mining, and these rules are converted into logical expressions to construct a constraint rule library.

[0082] in, This represents the number of rules. Each rule includes the trigger condition, the action to be performed, and the priority.

[0083] For the initial intervention sample set Each sample in the dataset is evaluated sequentially, including compliance assessment, authenticity assessment, and consistency assessment. Input the initial intervention sample set. (Covering typical risk scenarios but with quality differences), constraint rule base (Formalized policy rules and operational norms) Check whether the initial intervention sample meets the requirements. A compliance score is calculated for all hard constraints. A sample is invalidated if a hard constraint is violated, and a score is deducted for violating soft constraints.

[0084] (1) The formula for calculating the compliance score is:

[0085] in, The hard constraint weight is set to 0.7 in this embodiment. The soft constraint weight (set to 0.3 in this embodiment) is the weight coefficient set by the business; To impose strict constraints on the number of violations, For the number of violations of soft constraints, This represents the total number of constraints.

[0086] , These are business-defined, adjustable parameters used to flexibly adjust the emphasis on soft constraints in compliance assessments based on business needs. Hard constraints represent mandatory requirements that must be met (such as minimum standards in laws and regulations), and violations will result in non-compliance. Soft constraints are suggested or expected conditions (such as industry best practices or higher internal standards), and violations do not directly lead to non-compliance, but will affect the compliance score.

[0087] (2) Authenticity assessment calculation The similarity between candidate samples and real, manually corrected samples is measured as an assessment of authenticity. Edit distance is used to calculate the operational differences between candidate samples and real samples.

[0088] (3) Consistency assessment calculation The text is described by encoding operations using the BERT model. Cosine similarity is calculated (the semantic similarity between "correct income" and "adjust income" is 0.85). The intervention operation of the candidate sample is applied to the original decision trajectory. The consistency between the decision results before and after the intervention and the actual human correction results is compared. For example, if the actual result is "rejection" and the result after the intervention of the candidate sample is also "rejection", then the consistency score is 1.

[0089] The iterative optimization process is that the generator G is based on the current parameters. Combined with the initial input sample set and constraint rule base Generate candidate intervention samples .

[0090] like Violating the rules of sample evaluation calculation, such as triggering the hard constraint of "contribution base being lower than the minimum wage," will generate a negative reward. ,in The initial penalty coefficient for hard constraint violations is set as follows: ; If candidate intervention samples If all constraints are met (compliance score ≥ 0.9) and decision accuracy is improved (the decision changes from "incorrectly approved" to "correctly rejected" after intervention), then a positive reward is generated. ,in The initial reward coefficient for decision improvement is set at... . and The policy needs to be dynamically adjusted according to business scenarios, and the policy revision period is extended. (From 1.0 to 1.5) Strengthened penalties for violations of hard constraints, and increased penalties for batch business processing scenarios. (From 0.8 to 1.0) to encourage the generation of efficient samples that meet soft constraints.

[0091] The generator G adjusts its generation policy and updates its parameters based on the reward signal using the policy gradient algorithm. To maximize cumulative rewards:

[0092] Along gradient Direction update parameters The generator G learns to generate "high-reward" operations (i.e., operations that meet constraints and improve decision-making accuracy). The process stops when the proportion of compliant samples is ≥95% after five consecutive iterations, and the optimized intervention sample set is output. .

[0093] S4: Identify key nodes from the original decision trajectory of the robot process automation execution trajectory, obtain the comprehensive weight based on the adversarial contribution of the key nodes and the scenario-based interpretation weight matrix, and sum the basic interpretation vectors of the key nodes according to the comprehensive weight to obtain the enhanced interpretation vector. S401: Identify key nodes from the raw decision trajectory of the robotic process automation execution trajectory.

[0094] The system identifies the nodes (i.e., "critical nodes") that have the greatest impact on the final decision outcome from the original decision trajectory of the robotic process automation execution trajectory. Critical node identification provides a focus for subsequent "interpretive vector fusion" by quantifying the contribution of each node to the decision, ensuring that limited computational and interpretive resources are concentrated on the most critical aspects of the decision.

[0095] Key node identification requires combining data-driven contribution metric measurement with importance labeling based on business rules, comprehensively evaluating the "criticality" of nodes through multi-dimensional indicators. Data-driven contribution metric measurement relies on SHAP value calculation, based on cooperative game theory, and is applied to each node. The data-driven contribution metric is calculated, and the Kernel SHAP method is used to calculate the SHAP value of each node (reflecting the marginal contribution of a node to the decision outcome). The importance labeling of business rules is combined with the constraint rule base. This assigns business rule weights to nodes. By combining data-driven contribution with business rule weights, the overall criticality of a node is calculated. :

[0096] in, , The weighting coefficient (can be adjusted according to business needs; in this embodiment, it is set to...) , ), For nodes The SHAP value reflects the contribution at the data level. For nodes The weight of business rules reflects the importance of compliance / business objectives.

[0097] A threshold (CS ≥ 0.6) is set to filter key nodes. Nodes with SHAP values ​​exceeding the threshold are marked as key nodes. High-key nodes (CS ≥ 0.8) play a decisive role in the decision-making outcome. Medium-key nodes (0.6 ≤ CS < 0.8) and low-key nodes (CS < 0.6) are selected to form the set of key nodes. .

[0098] S402: Calculate the adversarial contribution of key nodes, obtain the comprehensive weight based on the adversarial contribution and the scenario-based interpretation weight matrix, identify the interpretation information of key nodes to obtain the basic interpretation vector, and strengthen the basic interpretation vector based on the comprehensive weight to obtain the enhanced interpretation vector.

[0099] The explanatory information of the identified key nodes (SHAP value, node type, business logic) is deeply integrated with the scenario-based explanatory weight matrix and adversarial contribution (from the perturbation module of I-EvGAN) to generate a reinforced explanatory vector. The core contribution of the key nodes is retained, and then the dynamic needs of the business scenario (policy weight, risk priority) and the impact of adversarial perturbation (deviation-prone nodes) are integrated to ultimately support business services such as visual report generation and manual review.

[0100] The adversarial contribution is calculated by extracting the output change of each key node under adversarial perturbation from the generator of I-EvGAN, and denoted as the adversarial contribution. :

[0101] in, The decision probability of the original sample. This represents the decision probability of the perturbation sample (the intervention sample generated by the generator). For example, the decision change probability of the "policy clause node" increases from 0.6 to 0.9 after the perturbation, contributing 0.3.

[0102] Calculate weighted fusion to assess adversarial contribution. With contextualized interpretation of the weight matrix corresponding dimension weights A joint decision is made, resulting in a comprehensive weight. :

[0103] in, This indicates a positive adjustment against the contribution level. The larger the value, the more sensitive the node is to disturbances, and the weight needs to be appropriately increased to indicate the risk. For the normalization term, ensure The total weight is 1, which facilitates weighted summation.

[0104] After identifying key nodes, a basic explanatory vector is generated for each node. The fundamental explanatory vectors of key nodes By overall weight Weighted summation generates enhanced interpretation vectors : .

[0105] Generating a basic explanatory vector for each node involves transforming node features into computable numerical vectors. Features include node type (operation type / decision type), timestamp, execution order, and associated modules (RPA / large model). If a key node is a rule-driven decision step, such as triggering an alert in RPA if the risk score > 80, an explanatory vector can be generated through rule extraction. The specific steps are as follows: (1) Rule parsing: Extract the conditions (risk score > 80) and actions (trigger warning) from the nodes; (2) Semantic encoding: Encode the rule conditions (numerical thresholds, logical relationships) and actions (operation types) into vectors (hot encoding of the "risk score" dimension and the "trigger warning" operation type); (3) Weighting: Assign weights to each dimension based on the importance of the rules in the process, and generate... .

[0106] S5: Evaluate the intervention samples based on the evaluation metrics to obtain the evaluation results; retrain the intervention generative adversarial evolution network based on the evaluation results to obtain the optimized intervention generative adversarial evolution network; obtain the optimized intervention samples based on the optimized intervention generative adversarial evolution network; S501: Define evaluation metrics.

[0107] Quantitative indicators are used to measure the intervention sample (intervention sample set). The quality of samples must be ensured to meet core requirements such as compliance, completeness, and ambiguity.

[0108] Compliance Score (CS): Checks whether the reinforcement explanatory vector covers all necessary provisions of the relevant policies, and is calculated with reference to the compliance score.

[0109] Completeness Score (CoS): This assesses whether the explanation includes all the key factors required for the decision. For example, "rejecting unemployment benefits" requires key factors such as "voluntary resignation" and "social security contributions suspended for less than 6 months." Completeness is calculated using coverage ratio.

[0110] Ambiguity Score (AS): Using text ambiguity detection models such as BERT-based semantic similarity classifiers, this score determines whether the interpretation contains multiple meanings. .

[0111] S502: Subsequent retraining of I-EvGAN. The typical training process of I-EvGAN includes initial adversarial training (unsupervised pre-training + adversarial training) and subsequent retraining (optimizing the model based on new data or task objectives).

[0112] By optimizing the generator's (such as generator G in I-EvGAN) sample generation strategy through targeted training based on evaluation metrics, issues such as insufficient compliance, lack of completeness, and excessive ambiguity identified in the S501 evaluation are addressed. The triggering condition is CS < 0.8, CoS < 0.7, or AS > 0.2, indicating an insufficient sample. Ultimately, a high-quality, compliant, and interpretable intervention sample set is generated. .

[0113] Constraint rule library The decoder layer, integrated into the generator, filters out illegal operations in real time; and dynamically adjusts the priority of constraint rules based on the current business scenario.

[0114] A multi-objective loss function is designed by combining evaluation metrics for compliance, completeness, and ambiguity:

[0115] Among them, satisfying .

[0116] Discriminator D training: Input real high-quality samples (from historical annotation databases) and samples generated by the generator. Output the probability of truth. And compliance / completeness / ambiguity score.

[0117] Generator G training: The generator attempts to generate samples that the discriminator cannot distinguish from real samples (maximizing...). ), while minimizing the total loss . Adjust the perturbation constraints of I-EvGAN by adding a regularization term to the generator that requires "key nodes to be modified" until the newly generated samples pass the evaluation. Output the optimized generator. Generative models with adjusted parameters, strategies, or structures, and high-quality sample sets. Intervention samples that meet quality requirements after retraining.

[0118] S6: Generate a decision report based on the enhanced explanatory vector and intervention samples, showcasing the decision intervention logic, decision process, and optimization results from multiple dimensions. Supports multi-dimensional display including time, features, and scenarios.

[0119] The complex decision-making logic and sample information are transformed into visual reports that business personnel can quickly understand, and then integrated with the manual review process.

[0120] Based on enhanced interpretation vectors Compared with the optimized intervention sample It generates structured and scenario-based decision reports, supporting multi-dimensional presentation of decision-making basis and process. It enhances explanatory vectors. Used for quantifying decision-making basis: Transforming the contribution of key nodes to the final decision into understandable numerical vectors, clarifying which nodes / features dominate the decision; Enhancing interpretability: Transforming abstract decision-making logic into explanatory language understandable to business personnel, for example, a weight reduction is triggered due to low matching degree of policy clause ID002. Optimized intervention sample. Used to provide corrective examples: showcasing the ideal corrected decision trajectory (adjusted weights, optimized RPA operation paths, etc.) as a reference benchmark for business personnel review; verifying the validity of the interpretation: verifying the effectiveness by using the actual intervention effect of the sample (whether the corrected decision is approved). The accuracy of the explanation logic, such as whether the adjustment of the explanation weights actually improved the decision-making); scenario-based demonstrations: samples covering typical risk scenarios (policy change period, high-risk areas) to provide business personnel with decision-making correction examples in multiple scenarios.

[0121] The decision report generated based on the reinforcement explanatory vector and intervention samples is as follows: (1) Data alignment: Explanation of each dimension (policy clause matching degree, risk score impact, etc.) and Intervention samples (adjusted weight values, corrected RPA trajectories, etc.) are aligned by timestamp or decision node.

[0122] (2) Extraction of key information: from Extract high-contribution nodes (such as "Policy Clause ID002") and Key interventions in the process (such as "increasing the weight from 0.3 to 0.5").

[0123] (3) Logical connection: Based on time series or decision-making process, the explanation and intervention operation are connected into a narrative chain of "decision background → key node analysis → intervention measures → expected effect".

[0124] Based on the above generation steps, the core content of the decision report is shown in the following example: Decision background: Description of the original decision scenario (e.g., "A user applies for pension certification, and the system's initial decision is 'reject'"), reasons for triggering correction (e.g., "Manual review found low policy clause matching"); Key node analysis: Based on List the nodes with the greatest impact on decision-making (such as "policy clause matching degree" and "risk score"), and quantify their contribution (such as "policy clause matching degree contribution -0.4, leading to initial rejection"); Intervention measures and effects: Show Intervention samples (e.g., "adjusting the policy clause matching weight to 0.6") and comparing the decision results before and after the intervention (e.g., "the revised decision is 'pass'"); Risk warnings and suggestions: based on resistance contribution High-risk nodes are identified (e.g., "Decisions are easily reversed after a node is disturbed, and manual review needs to be strengthened"), and optimization suggestions are proposed (e.g., "Increase the monitoring frequency of this node").

[0125] By highlighting key points and providing risk warnings in the report, manual reviewers are guided to focus on high-risk areas, highly sensitive points, and samples with questionable compliance, thus enabling efficient manual review and improving review efficiency.

[0126] The specific steps for manual review are as follows: (1) Explanation of logical verification Verify key nodes: Check whether the key nodes listed in the report are consistent with the original decision trajectory (e.g., whether "policy clause ID002 was indeed ignored in the original trajectory"). Verify contribution: Confirm the reasonableness of the node's contribution to the decision through business experience (e.g., whether "policy clause matching degree -0.4 is consistent with the actual impact").

[0127] (2) Assessment of the reasonableness of the intervention sample Matching business needs: Inspection Whether the intervention sample in the study aligns with business objectives (e.g., whether "increasing the weight to 0.5 is consistent with the company's risk appetite"); Simulated execution results: The intervention sample (e.g., "applying the adjusted weights") is simulated using RPA or a large model to verify whether it can effectively correct the original decision (e.g., "whether the correction passes certification").

[0128] 3. Report completeness and traceability check Information completeness: Confirm whether the report covers core modules such as decision-making background, key milestones, intervention measures, and risk warnings; Traceability: Check the correlation between the intervention sample and the original decision trajectory (e.g., whether the intervention sample was generated based on real RPA operation records) to ensure that the report can be traced back to specific business operations.

[0129] 4. Feedback and Iteration Problem Log: If any logical deviations in the explanation are found (such as "omission of key nodes") or unreasonable intervention samples are found (such as "excessive weight adjustment"), record the problem and report it to the S4 and S5 modules for optimization; Knowledge Accumulation: High-quality reports that pass review are stored in a knowledge base to provide a reference for decision-making corrections in similar scenarios in the future.

[0130] The corrections made by manual reviewers are fed back into the evolutionary optimization process, driving the generator to continuously optimize and serving as dynamic feedback on the decision-making results. Specifically, this feedback optimization involves: (1) Feedback optimization S3 1) Optimization process Training data selection for generator (G): If the feedback is "the generated sample is very different from the real intervention", optimize the input data of S3 and select high-quality samples that are closer to human correction (increase the proportion of real human correction samples or remove noisy samples) to provide the generator with a better "real sample" supervision signal.

[0131] Evolutionary algorithm parameter adjustment: If the feedback is "insufficient diversity of generated samples", adjust the parameters of the evolutionary algorithm or introduce an adaptive strategy to improve the diversity and adaptability of generated samples.

[0132] Loss function redesign: If the feedback is "the impact of the generated samples on the model decision does not meet expectations", optimize the generator's loss function and add an evaluation item "the effect of intervention samples on the correction of model decision" (such as "the accuracy of the corrected decision is improved") to replace the single "realism loss" and guide the generator to generate more practical samples.

[0133] 2) Optimize data Add real human correction samples: Collect "high-value correction samples" from human review and use them as training data for S3 to enhance the generator's learning of effective interventions.

[0134] Adversarial perturbation data for labeled samples: In response to feedback that "generated samples have weak anti-perturbation ability", adversarial perturbation samples are generated (slightly modifying the features of the original samples) and added to S3 training data to improve the robustness of the generator to perturbations.

[0135] (2) Feedback optimization S4: 1) Optimization process Improvements to the key node identification algorithm: If feedback indicates "key nodes are missing or mislabeled", optimize the key node identification logic in S4, introduce a more accurate contribution measurement method, or combine business rules to assist in identification.

[0136] The explanatory vector generation method has been upgraded: if the feedback is "the explanatory vector cannot reflect the actual decision-making logic", the basic explanatory vector will be optimized. The generation method employs multimodal fusion, etc.

[0137] Fusion strategy optimization: If the feedback is "enhanced explanation vector" "Information redundancy or lack of focus" requires adjustment of the weighting method of the basic explanatory vector (dynamic allocation of node weights), or addition of "business relevance" filtering (retaining only high-contribution nodes relevant to the current decision-making scenario).

[0138] 2) Optimize data Manually labeled explanatory data: "Correctly interpreted samples" (such as "the main reason for the decision was the improved matching degree of policy clauses") are collected from manual review and used as supervision signals to train the S4 explanatory vector generation model, thereby improving the accuracy of the explanation.

[0139] Domain Knowledge Graph: Construct a business domain knowledge graph, input S4 as prior knowledge, to guide the identification of key nodes and the generation of interpretation vectors (such as automatically associating "policy clause ID002" with "payment years calculation").

[0140] In summary, this invention aims to address the challenges of intelligent decision review and interpretation in existing large-scale human resources and social security decision-making systems, such as insufficient interpretability, difficulties in manual review, error amplification, and compliance risks. It proposes a multi-dimensional adaptive and intervention-based evolutionary interpretation optimization method for large-scale human resources and social security decision-making. This method constructs multi-dimensional interpretive representations, designs an adaptive weight adjustment mechanism, and combines intervention-generated adversarial evolutionary networks with policy and rule constraints. This enables the identification, interpretation enhancement, and visual feedback of key nodes in the decision-making process, significantly improving the efficiency of manual review and the system's auditability.

[0141] Example 2 This embodiment discloses a decision intervention interpretation and optimization system based on a large-scale human resources and social security model, including: The interpretation representation construction module is configured to: acquire multi-source data from the human resources and social security business system, extract multi-dimensional heterogeneous features based on the multi-source data, vectorize the multi-dimensional heterogeneous features, and form a multi-dimensional interpretation representation space. The adaptive weight adjustment module is configured to: perform preliminary weight allocation on the features of each dimension in the multidimensional interpretation representation space, filter key data, input the key data into the adjustment model to dynamically adjust the preliminary weights, and obtain the updated scenario-based interpretation weight matrix. The intervention sample generation module is configured to: obtain the original decision trajectory of the large model reasoning path and the robot process automated execution trajectory, and input it together with the scenario-based interpretation weight matrix into the intervention generation adversarial evolution network for processing to generate intervention samples; The interpretation enhancement module is configured to: identify key nodes from the original decision trajectory of the robot process automation execution trajectory, obtain a comprehensive weight based on the adversarial contribution of the key nodes and the scenario-based interpretation weight matrix, and sum the basic interpretation vectors of the key nodes according to the comprehensive weight to obtain the enhanced interpretation vector; The decision report generation module is configured to generate a decision report based on the enhanced explanatory vector and intervention samples, and to display the decision intervention logic, decision process and optimization results from multiple dimensions.

[0142] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0143] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0144] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0145] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0146] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0147] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A decision intervention explanation optimization method based on human social big model decision-making, characterized in that, The method comprises the following steps: acquiring multi-source data of a human resources business system, extracting multi-dimensional heterogeneous features based on the multi-source data, vectorizing the multi-dimensional heterogeneous features, and forming a multi-dimensional explanatory representation space; performing preliminary weight distribution on each dimension feature in the multi-dimensional explanatory representation space, screening key data, inputting the key data into an adjustment model to dynamically adjust the preliminary weight, and obtaining an updated scenario-based explanatory weight matrix; acquiring a large model inference path and an original decision trajectory of a robotic process automation execution trajectory, inputting the scenario-based explanatory weight matrix into an intervention generation adversarial evolution network together with the original decision trajectory, and processing to generate an intervention sample; identifying a key node from the original decision trajectory of the robotic process automation execution trajectory, obtaining a comprehensive weight based on the adversarial contribution degree of the key node and the scenario-based explanatory weight matrix, weighting and summing the basic explanatory vectors of the key node according to the comprehensive weight, and obtaining a reinforced explanatory vector; generating a decision report based on the reinforced explanatory vector and the intervention sample, and multi-dimensionally displaying decision intervention logic, a decision process and optimization results.

2. The decision intervention explanation optimization method based on human social model decision according to claim 1, wherein, The multi-dimensional heterogeneous features include policy clause basis, risk factors, business logic rules, historical case matching degree and real-time data consistency.

3. The decision intervention explanation optimization method based on human social model decision according to claim 1, wherein, The key data includes a historical misjudgment case library, a policy change log and a real-time risk situation.

4. The decision intervention explanation optimization method based on human social model decision according to claim 1, wherein, The intervention generation adversarial evolution network includes a generator, a discriminator and an adversarial perturbation module, and the adversarial perturbation module guides the generator to preferentially perturb high-weight features according to the scenario-based explanatory weight matrix.

5. The decision intervention explanation optimization method based on human social model decision according to claim 4, characterized in that, The intervention sample is generated by: obtaining a trained generator and discriminator through two-stage training, generating a preliminary intervention sample based on the trained generator; evolutionally optimizing the preliminary intervention sample through policy rule constraints to obtain an optimized intervention sample.

6. The decision intervention explanation optimization method based on human social model decision according to claim 5, characterized in that, The policy rule constraints include mandatory hard constraints, guiding soft constraints, conditional constraints and priority constraints.

7. The decision intervention explanation optimization method based on human social model decision according to claim 5, characterized in that, Based on the evaluation index, the optimized intervention sample is evaluated to obtain an evaluation result; based on the evaluation result, the intervention generation adversarial evolution network is retrained to obtain an optimized intervention generation adversarial evolution network; based on the optimized intervention generation adversarial evolution network, an optimized intervention sample is obtained, and a decision report is generated based on the reinforced explanatory vector and the optimized intervention sample.

8. A decision intervention explanation optimization system based on human social big model decision-making, characterized in that, The method comprises the following steps: An explanatory representation construction module is configured to acquire multi-source data of a human resources business system, extract multi-dimensional heterogeneous features based on the multi-source data, vectorize the multi-dimensional heterogeneous features, and form a multi-dimensional explanatory representation space; An adaptive weight adjustment module is configured to perform preliminary weight distribution on each dimension feature in the multi-dimensional explanatory representation space, screen key data, input the key data into an adjustment model to dynamically adjust the preliminary weight, and obtain an updated scenario-based explanatory weight matrix; An intervention sample generation module is configured to acquire a large model inference path and an original decision trajectory of a robotic process automation execution trajectory, input the scenario-based explanatory weight matrix into an intervention generation adversarial evolution network together with the original decision trajectory, and process to generate an intervention sample; The explanation reinforcement module is configured to: identify key nodes from an original decision trajectory of a robot process automation execution trajectory, obtain a comprehensive weight based on an adversarial contribution degree of the key nodes and a scenario explanation weight matrix, weight and sum basic explanation vectors of the key nodes according to the comprehensive weight to obtain a reinforced explanation vector; The decision report generation module is configured to: generate a decision report based on the reinforced explanation vector and an intervention sample, and multi-dimensionally display decision intervention logic, a decision process and optimization results.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the decision intervention explanation optimization method based on human social big model decision according to any one of claims 1-7.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the decision intervention explanation optimization method based on human social big model decision according to any one of claims 1-7.

Images