Insurance service intelligent decision engine based on big data and business intelligence

Abstract

The invention relates to the technical field of intelligent decision-making, and discloses an insurance service intelligent decision-making engine based on big data and business intelligence, which comprises the steps of acquiring input data, performing confrontation training, comparing rule recognition accuracy before and after the confrontation training, constructing a dynamic rule base, and calculating a data sensitivity level. Constructing a compliance decision for responding to the supervision policy and the data sensitivity level in real time; a group Gini coefficient is calculated, the input data distribution balance condition is detected, a fairness weight matrix is constructed for real-time fairness feedback, a dynamic fairness constraint condition is constructed, a relaxation factor is adjusted, and a target function with fairness constraint is generated; constructing a virtual anti-fact sample, verifying the decision fairness of the decision model, calculating an anti-fact discrimination measure, and positioning a discrimination source; a decision path contribution decomposition algorithm is put forward, feature contribution adjustment values are calculated, feature contributions are fairly distributed, key features and decision nodes causing prejudice are positioned, and a fair perception reinforcement learning correction mechanism is established.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, specifically to an intelligent decision-making engine for insurance services based on big data and business intelligence. Background Technology

[0002] Traditional insurance decision-making relies on manual underwriting and experience-driven approaches, which suffers from low efficiency and inaccurate risk assessment. With the development of big data technology, insurance companies have accumulated customer data and external data, providing a data foundation for intelligent decision-making. BERT can process unstructured text, deep learning can model time-series data, and reinforcement learning can achieve dynamic decision optimization, providing technical support for intelligent decision engines.

[0003] In existing technologies, traditional actuarial science relies on the law of large numbers and static historical data, while AI models require real-time dynamic feedback. This leads to the instability of historical risk pools, a lack of historical data for extreme disasters, and difficulty in passing regulatory verification for virtual data synthesized by adversarial networks. Existing technologies largely rely on manual updates to regulatory rules, resulting in delayed policy responses. The rule base lacks a dynamic adjustment mechanism, making it difficult to cope with frequent policy changes and rule conflicts. Fairness assessments of insurance decisions often focus on single sensitive attributes, ignoring discrimination against overlapping groups. Fairness constraints are based on fixed thresholds, failing to balance decision accuracy and fairness. Insurance data is scattered across multiple entities such as insurance companies and medical institutions, making centralized modeling prone to privacy leaks. Traditional counterfactual analysis relies on centralized data, making cross-institutional implementation difficult. Existing technologies often use static labels to determine the level of sensitive information such as health data, ignoring the dynamic changes of data over time.

[0004] Therefore, there is a need to provide an intelligent decision-making engine for insurance services based on big data and business intelligence. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent decision-making engine for insurance services based on big data and business intelligence. To solve the aforementioned problems in the prior art, this invention achieves this through the following technical solution:

[0006] In a first aspect, the intelligent decision-making engine for insurance services based on big data and business intelligence provided in this embodiment of the invention specifically includes the following units:

[0007] Engine Adjustment Unit: Acquires input data, performs adversarial training based on the input data, compares the rule recognition accuracy before and after adversarial training, builds a dynamic rule base, monitors changes in data sensitivity level, calculates data sensitivity level, and builds compliance decisions that respond in real time to regulatory policies and data sensitivity levels;

[0008] Modeling and inference unit: Calculates the group Gini coefficient based on the input data of adversarial training, detects the balance of the input data distribution. If the distribution is uneven, a data augmentation mechanism is triggered. If the distribution is uniform, fair data is synthesized, a fairness weight matrix is ​​constructed for real-time fairness feedback, the group cross-fairness index is calculated to construct dynamic fairness constraints, and the relaxation factor is dynamically adjusted to generate an objective function with fairness constraints.

[0009] Verification and localization unit: Construct virtual counterfactual samples under the federated learning framework to verify the fairness of the decision-making model, calculate the counterfactual discrimination metric, and locate the source of discrimination caused by sensitive attributes in insurance decisions;

[0010] Fair Decision Unit: Proposes a decision path contribution decomposition algorithm to calculate feature contribution adjustment values ​​and fairly allocate feature contributions. Locates key features and decision nodes that lead to bias in insurance decisions, establishes a fairness perception reinforcement learning correction mechanism, and achieves closed-loop optimization of bias correction.

[0011] Secondly, the intelligent decision-making method for insurance services based on big data and business intelligence provided in this embodiment of the invention specifically includes the following steps:

[0012] Step 1: Acquire input data, conduct adversarial training based on the input data, compare the rule recognition accuracy before and after adversarial training, build a dynamic rule base, monitor changes in data sensitivity level, calculate the data sensitivity level, and build a compliance decision system that responds in real time to regulatory policies and data sensitivity levels.

[0013] Step 2: Calculate the group Gini coefficient based on the input data of adversarial training, detect the balance of the input data distribution. If the distribution is uneven, trigger the data augmentation mechanism. If the distribution is uniform, synthesize fairness data, construct a fairness weight matrix for real-time fairness feedback, calculate the group cross fairness index to construct dynamic fairness constraints, dynamically adjust the relaxation factor, and generate an objective function with fairness constraints.

[0014] Step 3: Construct virtual counterfactual samples within the federated learning framework to verify the fairness of the decision-making model, calculate the counterfactual discrimination metric, and pinpoint the sources of discrimination in insurance decisions caused by sensitive attributes;

[0015] Step 4: Propose a decision path contribution decomposition algorithm to calculate the feature contribution adjustment value and fairly allocate feature contributions. Locate the key features and decision nodes that lead to bias in insurance decision-making, establish a fairness perception reinforcement learning correction mechanism, and achieve closed-loop optimization of bias correction.

[0016] The beneficial effects of this invention are:

[0017] 1. Adversarial training is introduced into insurance regulatory policy rule recognition. Adversarial samples are generated by adding small perturbations to the word vectors of policy texts, enhancing the model's tolerance to semantic perturbations such as synonym substitution and fine-tuning of expression, and solving the problem of insufficient robustness of traditional rule recognition models to policy text variants. A real-time policy acquisition mechanism is constructed, and triple rule extraction forms a dynamic rule base and introduces a priority mechanism to achieve real-time response to regulatory policies and intelligent resolution of conflicting rules. A two-layer fairness measurement framework is proposed. The group Gini coefficient detects the data distribution balance, and the group cross-fairness index quantifies the potential discrimination of multi-dimensional cross groups, breaking through the limitations of traditional single-dimensional fairness assessment. A fairness constraint objective function with a dynamic relaxation factor is designed, and the constraint strength is adjusted in real time according to the change rate of the group cross-fairness index to balance decision accuracy and fairness.

[0018] 2. Virtual counterfactual samples are generated, and differential privacy protection is combined to achieve discrimination detection in cross-institutional data silos, addressing the privacy leakage risks of traditional centralized data processing. A counterfactual discrimination metric is proposed, which accurately locates the source of discrimination through statistical significance testing, providing a quantitative basis for fairness correction. An actual feature contribution adjustment algorithm fairly allocates the contribution of features to decision-making, generating bias heatmaps and decision path maps to achieve visualized location of discriminatory features. A fairness-aware reinforcement learning correction mechanism is constructed, using counterfactual discrimination metric, prediction accuracy, and privacy risk as the state space, and optimizing decision-making by adjusting feature weights and decision threshold actions to form a closed loop. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the steps of the intelligent decision-making engine for insurance services based on big data and business intelligence provided in Embodiment 1 of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the insurance service intelligent decision engine based on big data and business intelligence provided in Embodiment 2 of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] Example 1: As Figure 1 As shown in the figure, the intelligent decision-making engine for insurance services based on big data and business intelligence provided in this embodiment of the invention specifically includes the following units:

[0024] Engine Adjustment Unit: Acquires input data, performs adversarial training based on the input data, compares the rule recognition accuracy before and after adversarial training, builds a dynamic rule base, monitors changes in data sensitivity level, calculates data sensitivity level, and builds compliance decisions that respond in real time to regulatory policies and data sensitivity levels;

[0025] In a specific embodiment, policy documents published on the official websites of the State Financial Regulatory Commission and the National Health Commission are crawled in real time using web crawlers.

[0026] Build a distributed web crawler cluster to dynamically crawl web page content, and set up an incremental crawling mechanism for official websites that are updated frequently.

[0027] Access regulatory agencies' open APIs as a supplementary data source for web crawlers to ensure the completeness of policy information.

[0028] Using the BERT model for text classification and entity recognition, we extracted key compliance requirements for cross-border transmission of health data that require separate authorization, and generated a structured rule base.

[0029] Manually labeled policy text samples were used to train a BERT model to perform two types of tasks:

[0030] Text classification: Identifying whether policies involve pre-defined core areas;

[0031] Entity identification: Extracting key constraints, responsible parties, and timeliness requirements from policies;

[0032] Transform unstructured text into triple rules, including: subject, behavior, and constraint;

[0033] The adversarial training mechanism FGSM is used to enhance the robustness of the decision model to semantic perturbations of policy texts and improve the accuracy of rule recognition.

[0034] Acquire input data and perform adversarial training based on the input data;

[0035] It should be noted that the input data includes both structured and unstructured data. Structured data includes, but is not limited to, customer income levels, residential area safety index, medical record codes, age, occupation type, past claims records, and policy type. Unstructured data includes, but is not limited to, insurance application text descriptions, customer feedback texts, and doctor's diagnosis descriptions. Keywords in the insurance application text descriptions are extracted using NLP; for example, keywords include: pre-existing conditions and high-risk occupations.

[0036] During the training of the decision model, small perturbations are added to the word vectors of the input text to generate adversarial examples;

[0037] The decision-making model enhances its tolerance to semantic perturbations such as synonym substitution and wording fine-tuning in policy texts by learning adversarial examples.

[0038] By comparing the rule recognition accuracy before and after adversarial training, we can determine whether the preset accuracy rate is achieved.

[0039] If the preset accuracy rate is not achieved, continue adversarial training;

[0040] If the preset accuracy rate is achieved, a dynamic rule base is constructed based on the parsed triple rules. The rule attributes include, but are not limited to: rule ID, effective time, applicable scenario, constraints, and violation penalty measures.

[0041] Set a rule priority mechanism to automatically trigger the higher priority rule when multiple rules conflict;

[0042] For example, the priority mechanism is set as follows: legal provisions > departmental regulations > industry guidelines;

[0043] Real-time collection of customer health data, including medical examination reports and medical insurance settlement records;

[0044] Construct an LSTM-based time-series classification model to monitor changes in data sensitivity levels using the formula:

[0045] S t =σ s (W s *[h t-1 ,x t ]+b s )

[0046] The data sensitivity level S at time t is calculated. t , where h t-1 Let x be the hidden state at time t-1. t Let σ be the current health indicator vector. s (W s *[h t-1 ,x t ]+b s) is the Sigmoid activation function, W s The weight matrix calculated for the sensitivity level, b s The bias term is calculated for the sensitivity level, and the attention mechanism is used to capture the temporal dependencies between health indicators;

[0047] Based on the policy rule base and data sensitivity level, compliance decisions are triggered in real time, and data streams such as medical insurance catalog updates and meteorological disaster early warnings are obtained;

[0048] Design an event-driven rules engine that automatically calls the compliance rules library and generates risk handling instructions when a cross-border health data transfer event is triggered.

[0049] The rules engine loads triple rules from the policy rule library and automatically matches relevant rules when an event is triggered.

[0050] It should be noted that GraphQL technology is used to achieve real-time subscription and intelligent routing of multi-source data streams, reducing decision latency;

[0051] Modeling and inference unit: Calculates the group Gini coefficient based on the input data of adversarial training, detects the balance of the input data distribution. If the distribution is uneven, a data augmentation mechanism is triggered. If the distribution is uniform, fair data is synthesized, a fairness weight matrix is ​​constructed for real-time fairness feedback, the group cross-fairness index is calculated to construct dynamic fairness constraints, and the relaxation factor is dynamically adjusted to generate an objective function with fairness constraints.

[0052] The population Gini coefficient is calculated based on the input data from adversarial training to detect the equilibrium of the input data distribution, using the formula:

[0053]

[0054] The population Gini coefficient G of the k-th group was calculated. k , where x i x j These are all feature values ​​within the population, where n is the population sample size. The mean of the group characteristics;

[0055] The obtained population Gini coefficient is compared with the preset minimum population Gini coefficient. If the population Gini coefficient is greater than the preset minimum population Gini coefficient, it indicates that the population data is unevenly distributed, triggering the data augmentation mechanism.

[0056] If the group Gini coefficient is less than or equal to the preset minimum group Gini coefficient, then fair data is synthesized to force the generated data distribution of different groups to be consistent, so as to avoid the synthesized data from reinforcing the original bias.

[0057] Generator loss function design:

[0058] LG =E[D(G(z|c))]+λ L *FairLoss(c)

[0059] The generator loss L was calculated. G Where E is the expectation function of the probability distribution of the input, D is the discriminator model function, G is the generator model function, z is the random noise vector, c is the population condition, and λ L The fairness loss weight is preset to an empirical value of 0.8, and FairLoss(c) is the fairness loss term;

[0060] The fairness loss item is calculated using the following formula:

[0061] FairLoss(c) = ||E[D(G(z|c)] i ))]-E[D(G(z|c j ))]||2

[0062] The fairness loss term FairLoss(c) is calculated, where c i c j All are group conditions;

[0063] A fairness weight matrix W∈R is constructed based on group size, data quality, and fairness requirements. m×n , where m is the population size and n is the feature dimension;

[0064] Real-time fairness feedback is provided based on a fairness weight matrix, and the strength of fairness constraints is dynamically adjusted according to the actual impact on insurance decisions. The cross-group fairness index is calculated to quantify potential discrimination between multi-dimensional cross-groups, which is suitable for complex insurance scenarios.

[0065] Based on real-time fairness feedback, the cross-fairness index of the groups is calculated using the formula:

[0066]

[0067] The cross-equity index C of the group was calculated. FID Where H is the total number of feature dimensions, w h The dimension weight of the h-th dimension is calculated using the entropy weight method. The Kolmogorov-Sminrnov statistic measures the difference in distribution between population A and population B along the h-th dimension. Let be the cumulative distribution function of group A and group B on the d-th dimension;

[0068] Dimension weights are calculated using the entropy weighting method:

[0069]

[0070] The calculation yields, where e d Let p be the information entropy of the d-th dimension. id Let n be the probability density of the i-th sample in the d-th dimension, n be the number of samples, and i be the sample index.

[0071] Based on the calculated group cross-fairness index, dynamic fairness constraints are constructed, using the core constraint formula:

[0072]

[0073] Implement dynamic fairness constraints, among which... For the prediction results of group A and group B, P A P B δ represents the proportions of group A and group B in the total sample. t C is the time-varying relaxation factor, t is the timestamp, and C is the time-varying relaxation factor. FID (A,B) represents the cross-equity index between groups A and B;

[0074] Based on the obtained group cross-fairness index, monitor the rate of change of the group cross-fairness index in the decision output in real time;

[0075] The relaxation factor is dynamically adjusted based on the rate of change of the group cross-equity index.

[0076] If the rate of change of the group cross-fairness index exceeds the preset threshold for the rate of change of the index (empirically set to 0.15), the constraint reinforcement is activated, and the time-varying relaxation factor δ is reduced. t Tighten the boundaries of fairness;

[0077] If the rate of change of the group cross-equity index is less than the preset threshold for the rate of change of the index, then the time-varying relaxation factor δ is relaxed. t Improve the accuracy of decision-making models;

[0078] Construct an objective function with fairness constraints:

[0079] min θ L(θ)+λ θ *max(0,V io (θ)-δ t *C FID )

[0080] Where L(θ) is the model loss function, θ is the model parameter, and λ is the model parameter. θ V is the constraint penalty coefficient. io (θ) represents the degree of constraint violation;

[0081] Verification and localization unit: Construct virtual counterfactual samples under the federated learning framework to verify the fairness of the decision-making model, calculate the counterfactual discrimination metric, and locate the source of discrimination caused by sensitive attributes in insurance decisions;

[0082] The FedAvg framework, which adopts the federal average, includes:

[0083] Central server: Coordinates all participants and aggregates model parameters.

[0084] Local nodes: Insurance company branches, partner medical institutions, etc., store local data and perform local training.

[0085] Encrypted communication layer: Homomorphic encryption is used to ensure secure parameter transmission;

[0086] Each participant generates virtual counterfactual samples locally; user non-sensitive characteristics Y are preserved. i No change, only the sensitive attribute S is modified. i ;

[0087] Add differential privacy protection:

[0088] in, The feature vector after adding noise is represented by Laplace(0,b), where Laplace(0,b) is the Laplace noise and b is the noise scale, set according to the privacy budget.

[0089] Based on differential privacy protection, the counterfactual discrimination metric is calculated using the formula:

[0090]

[0091] The counterfactual discrimination measure C was calculated. FD Where N is the sample size, f is the insurance decision model, S is the sensitive attribute, and k and m are different values ​​of the sensitive attribute, such as different income levels. To change the sensitive attribute of the i-th sample to k, the counterfactual sample is... The counterfactual sample is the sample whose sensitive attribute is changed to m.

[0092] Establish discrimination determination rules and use counterfactual discrimination measures to determine whether counterfactual discrimination exists;

[0093] Specifically, if the counterfactual discrimination measure is greater than the preset discrimination measure threshold and the probability of it occurring is greater than 0.01, i.e. the probability of significance is high, then statistically significant counterfactual discrimination is determined to exist.

[0094] The preset discrimination measurement threshold is set according to the type of insurance product. Each local node calculates the local CFD value and uploads it to the central server using a secure aggregation protocol. The central server then calculates the global CFD.

[0095] Fair Decision Unit: Proposes a decision path contribution decomposition algorithm to calculate feature contribution adjustment values ​​and fairly allocate feature contributions. Locates key features and decision nodes that lead to bias in insurance decisions, establishes a fairness perception reinforcement learning correction mechanism, and achieves closed-loop optimization of bias correction.

[0096] The fairness-perceived feature contribution is calculated using the Shapley Additive Explanation (SHAP) value, according to the formula:

[0097]

[0098] The characteristic contribution adjustment value was calculated. in, The feature contribution adjustment value φ is given to feature i after fair adjustment. i The original SHAP value is given, S is the set of sensitive features, and |S| is the number of sensitive features.

[0099] It should be noted that the Shapley additive explanatory value is an indicator that quantifies the contribution of each feature to the prediction result, and solves the problem of fairly allocating feature contributions.

[0100] Generate a bias heatmap, mark feature nodes that contribute more than a preset proportion to discrimination, construct a decision path graph, and highlight key decision branches that lead to group differences;

[0101] Output the fairness impact ranking of each feature, and quantify the counterfactual discrimination measure C of each feature on the overall population. FD The proportion of contribution;

[0102] Establish a fairness perception reinforcement learning correction mechanism;

[0103] The state space is: the current counterfactual discrimination measure C. FD Model prediction accuracy, group coverage, and privacy risk level calculated based on differential privacy parameters for each group;

[0104] The action space includes: adjusting feature weights, increasing the weights of non-sensitive features, reducing the influence of sensitive features, modifying the decision model structure, updating the data sampling strategy, and adjusting the decision threshold.

[0105] Obtain the model's prediction accuracy and privacy risk index, combine them with a counterfactual discrimination metric, and normalize them to construct a reward function: R = A cc -C FD1 -P R Among them, A cc To normalize the prediction accuracy of the model, C FD1 To normalize antifactual discrimination measures, P R To normalize the privacy risk index;

[0106] Reward signals are generated based on reward functions, decision execution results are collected, and reward signals of the fairness-aware reinforcement learning correction mechanism are updated.

[0107] The system assesses insurance risks based on a fairness constraint model, generates underwriting recommendations that include fairness scores, and automatically triggers manual review for high-risk cases with low fairness scores.

[0108] The system calculates the base premium based on a fairness perception model, provides personalized premium adjustment suggestions, and explains the premium composition.

[0109] Example 2

[0110] like Figure 2 As shown in the figure, the intelligent decision-making method for insurance services based on big data and business intelligence provided in this embodiment of the invention specifically includes the following steps:

[0111] Step 1: Acquire input data, conduct adversarial training based on the input data, compare the rule recognition accuracy before and after adversarial training, build a dynamic rule base, monitor changes in data sensitivity level, calculate the data sensitivity level, and build a compliance decision system that responds in real time to regulatory policies and data sensitivity levels.

[0112] Step 2: Calculate the group Gini coefficient based on the input data of adversarial training, detect the balance of the input data distribution. If the distribution is uneven, trigger the data augmentation mechanism. If the distribution is uniform, synthesize fairness data, construct a fairness weight matrix for real-time fairness feedback, calculate the group cross fairness index to construct dynamic fairness constraints, dynamically adjust the relaxation factor, and generate an objective function with fairness constraints.

[0113] Step 3: Construct virtual counterfactual samples within the federated learning framework to verify the fairness of the decision-making model, calculate the counterfactual discrimination metric, and pinpoint the sources of discrimination in insurance decisions caused by sensitive attributes;

[0114] Step 4: Propose a decision path contribution decomposition algorithm to calculate the feature contribution adjustment value and fairly allocate feature contributions. Locate the key features and decision nodes that lead to bias in insurance decision-making, establish a fairness perception reinforcement learning correction mechanism, and achieve closed-loop optimization of bias correction.

[0115] The above provides a detailed description of one embodiment of the present invention, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. The above formulas are all dimensionless numerical calculations, and the formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world situation. The preset parameters in the formulas are set by those skilled in the art based on actual conditions and historical experience, and can be adjusted according to actual conditions. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. An intelligent decision-making engine for insurance services based on big data and business intelligence, characterized in that: Includes the following units: Engine Adjustment Unit: Acquires input data, performs adversarial training based on the input data, compares the rule recognition accuracy before and after adversarial training, builds a dynamic rule base, monitors changes in data sensitivity level, calculates data sensitivity level, and builds compliance decisions that respond in real time to regulatory policies and data sensitivity levels; Modeling and inference unit: Calculates the group Gini coefficient based on the input data of adversarial training, detects the balance of the input data distribution. If the distribution is uneven, a data augmentation mechanism is triggered. If the distribution is uniform, fair data is synthesized, a fairness weight matrix is ​​constructed for real-time fairness feedback, the group cross-fairness index is calculated to construct dynamic fairness constraints, and the relaxation factor is dynamically adjusted to generate an objective function with fairness constraints. Verification and localization unit: Construct virtual counterfactual samples under the federated learning framework to verify the fairness of the decision-making model, calculate the counterfactual discrimination metric, and locate the source of discrimination caused by sensitive attributes in insurance decisions; Fair Decision Unit: Proposes a decision path contribution decomposition algorithm to calculate feature contribution adjustment values ​​and fairly allocate feature contributions. Locates key features and decision nodes that lead to bias in insurance decisions, establishes a fairness perception reinforcement learning correction mechanism, and achieves closed-loop optimization of bias correction.

2. The insurance service intelligent decision-making engine based on big data and business intelligence as described in claim 1, characterized in that, The method for conducting adversarial training is as follows: The adversarial training mechanism FGSM is used to enhance the robustness of the decision model to semantic perturbations of policy texts; Acquire input data and perform adversarial training based on the input data; During the training of the decision model, small perturbations are added to the word vectors of the input text to generate adversarial examples; The decision-making model enhances its tolerance to semantic perturbations such as synonym substitution and wording fine-tuning in policy texts by learning adversarial examples. By comparing the rule recognition accuracy before and after adversarial training, we can determine whether the preset accuracy rate is achieved. If the preset accuracy rate is not achieved, continue adversarial training; If the preset accuracy rate is achieved, a dynamic rule base is constructed based on the parsed triple rules. The rule attributes include, but are not limited to: rule ID, effective time, applicable scenario, constraints, and violation penalty measures. Set a rule priority mechanism to automatically trigger the higher-priority rule when multiple rules conflict.

3. The insurance service intelligent decision-making engine based on big data and business intelligence as described in claim 1, characterized in that, The method for obtaining the data sensitivity level is as follows: Real-time collection of customer health data, including medical examination reports and medical insurance settlement records; Construct an LSTM-based time-series classification model to monitor changes in data sensitivity levels, using the formula: S t =σ s (W s *[h t-1 ,x t ]+b s The data sensitivity level S at time t is calculated. t , where h t-1 Let x be the hidden state at time t-1. t Let σ be the current health indicator vector. s (W s *[h t-1 ,x t ]+b s ) is the Sigmoid activation function, W s The weight matrix calculated for the sensitivity level, b s The bias term is calculated for the sensitivity level, and the attention mechanism is used to capture the temporal dependencies between health indicators; Based on the policy rule base and data sensitivity level, compliance decisions are triggered in real time, and data streams such as medical insurance catalog updates and meteorological disaster early warnings are obtained; Design an event-driven rules engine that automatically calls the compliance rules library and generates risk handling instructions when a cross-border health data transfer event is triggered. The rules engine loads triple rules from the policy rule library and automatically matches the relevant rules when an event is triggered.

4. The insurance service intelligent decision-making engine based on big data and business intelligence as described in claim 1, characterized in that, The method for obtaining the population Gini coefficient is as follows: The population Gini coefficient is calculated based on the input data from adversarial training to detect the equilibrium of the input data distribution, using the formula: The population Gini coefficient G of the k-th group was calculated. k , where x i x j These are all feature values ​​within the population, where n is the population sample size. The mean of the group characteristics.

5. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 4, characterized in that, The method for obtaining the data distribution balance is as follows: The obtained population Gini coefficient is compared with the preset minimum population Gini coefficient. If the population Gini coefficient is greater than the preset minimum population Gini coefficient, it indicates that the population data is unevenly distributed, triggering the data augmentation mechanism. If the group Gini coefficient is less than or equal to the preset minimum group Gini coefficient, then the fairness data is synthesized.

6. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 1, characterized in that, The method for synthesizing fairness data is as follows: Force different groups to maintain a consistent distribution of generated data; Generator loss function design: L G =E[D(G(z|c))]+λ L *FairLoss(c) calculates the generator loss L. G Where E is the expectation function of the probability distribution of the input, D is the discriminator model function, G is the generator model function, z is the random noise vector, c is the population condition, and λ L FairLoss(c) is the fairness loss weight, and FairLoss(c) is the fairness loss term. The fairness loss term is calculated using the formula: FairLoss(c)=||E[D(G(z|c)] i ))]-E[D(G(z|c j The fairness loss term FairLoss(c) is calculated, where c i c j All of these are group conditions.

7. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 1, characterized in that, The method for obtaining the group cross-fairness index is as follows: A fairness weight matrix W∈R is constructed based on group size, data quality, and fairness requirements. m×n , where m is the population size and n is the feature dimension; Real-time fairness feedback is provided based on a fairness weight matrix, and the strength of fairness constraints is dynamically adjusted according to the actual impact on insurance decisions. The cross-group fairness index is calculated to quantify potential discrimination between multi-dimensional cross-groups, which is suitable for complex insurance scenarios. Based on real-time fairness feedback, the cross-fairness index of the groups is calculated using the formula: The cross-equity index C of the group was calculated. FID Where H is the total number of feature dimensions, w h The dimension weight of the h-th dimension is calculated using the entropy weight method. The Kolmogorov-Smirnov statistic measures the difference in distribution between population A and population B along the h-th dimension. Let be the cumulative distribution function of group A and group B on the d-th dimension.

8. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 7, characterized in that, The method for generating the objective function with fairness constraints is as follows: Based on the calculated group cross-fairness index, dynamic fairness constraints are constructed, using the core constraint formula: Implement dynamic fairness constraints, among which... For the prediction results of group A and group B, P A P B δ represents the proportions of group A and group B in the total sample. t C is a time-varying relaxation factor. FID C represents the cross-fairness index of the groups, where t is the timestamp and C is the cross-fairness index of the groups. FID (A,B) represents the cross-equity index between groups A and B; Based on the obtained group cross-fairness index, monitor the rate of change of the group cross-fairness index in the decision output in real time; The relaxation factor is dynamically adjusted based on the rate of change of the group cross-equity index. If the rate of change of the group cross-fairness index exceeds the preset threshold for the rate of change of the index, the constraint reinforcement is activated, and the time-varying relaxation factor δ is reduced. t Tighten the boundaries of fairness; If the rate of change of the group cross-equity index is less than the preset threshold for the rate of change of the index, then the time-varying relaxation factor δ is relaxed. t Improve the accuracy of decision-making models; Construct the objective function with fairness constraints: min θ L(θ)+λ θ *max(0,V io (θ)-δ t *C FID Where L(θ) is the model loss function, θ is the model parameter, and λ is the model parameter. θ V is the constraint penalty coefficient. io (θ) represents the degree of constraint violation, C FID This is the cross-equity index for the group.

9. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 1, characterized in that, The method for calculating the counterfactual discrimination metric is as follows: Each participant generates virtual counterfactual samples locally; user non-sensitive characteristics X are preserved. i No change, only the sensitive attribute S is modified. i ; Add differential privacy protection: in, The feature vector after adding noise is represented by Laplace(0,b), where Laplace(0,b) is the Laplace noise and b is the noise scale, set according to the privacy budget. Based on differential privacy protection, the counterfactual discrimination metric is calculated using the formula: The counterfactual discrimination measure C was calculated. FD Where N is the sample size, f is the insurance decision model, S is the sensitive attribute, and k and m are different values ​​of the sensitive attribute, such as different income levels. To change the sensitive attribute of the i-th sample to k, the counterfactual sample is... The counterfactual sample is the sample whose sensitive attribute is changed to m. Establish discrimination determination rules and use counterfactual discrimination measures to determine whether counterfactual discrimination exists.

10. The insurance service intelligent decision-making engine based on big data and business intelligence according to claim 1, characterized in that, The method for calculating the feature contribution adjustment value is as follows: The fairness-perceived feature contribution is calculated using the Shapley Additive Explanation (SHAP) value, according to the formula: The characteristic contribution adjustment value was calculated. in, The feature contribution adjustment value φ is given to feature i after fair adjustment. i The original SHAP value is given, S is the set of sensitive features, and |S| is the number of sensitive features. Generate a bias heatmap, mark feature nodes that contribute more than a preset proportion to discrimination, construct a decision path graph, and highlight key decision branches that lead to group differences; Output the fairness impact ranking of each feature, and quantify the counterfactual discrimination measure C of each feature on the overall population. FD The proportion of contribution; Establish a fairness perception reinforcement learning correction mechanism; The state space is: the current counterfactual discrimination measure C. FD Model prediction accuracy, group coverage, and privacy risk level calculated based on differential privacy parameters for each group; The action space includes: adjusting feature weights, increasing the weights of insensitive features, reducing the influence of sensitive features, modifying the decision model structure, updating the data sampling strategy, and adjusting the decision threshold.

Images