Artificial intelligence-based insurance customer fraud detection method and system

Abstract

The invention provides an insurance customer fraud detection method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: building a nonlinear mapping relation between a feature variable and a target variable, calculating a partial correlation coefficient between features, recognizing a causal relation, and generating an insurance fraud field causal atlas; identifying a trigger factor based on a time sequence intervention effect and an anti-fact effect; calculating a transfer coefficient and loop strength of the causal path to obtain a link credibility score; and fusing the trigger factor and the link credibility score to generate an early warning score, and extracting a maximum cumulative transfer effect path to output an early warning result. According to the method, the insurance fraud behavior can be accurately identified, and the accuracy and interpretability of fraud detection are improved.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to an insurance customer fraud detection method and system based on artificial intelligence. Background Art

[0002] Insurance fraud behaviors are increasing and the means are constantly being renovated. Traditional insurance fraud detection mainly relies on manual experience judgment and rule matching. However, with the growth of data volume and the complexity of fraud means, this method has been difficult to meet the actual needs. In recent years, with the development of artificial intelligence technology, fraud detection methods based on machine learning have gradually been applied. However, the existing insurance fraud detection technologies have the following deficiencies: Most of the existing technologies adopt the method of static feature modeling, only focusing on the surface correlation of data, ignoring the deep causal relationship behind fraud behaviors, and it is difficult to accurately grasp the evolution law and conduction mechanism of fraud behaviors, resulting in insufficient recognition ability for new fraud means.

[0003] When the existing methods conduct risk assessment, they often consider each risk factor separately in isolation, failing to fully consider the dynamic interaction and cumulative effect among risk factors, which easily causes the risk assessment results to be one-sided or distorted, and cannot effectively warn of potential systemic fraud risks.

[0004] The existing fraud detection results lack interpretability and evidence support, and it is difficult to clearly show the basis and reasoning process of fraud determination to business personnel, affecting the credibility and practical application effect of the detection results, and also being unfavorable for subsequent manual verification and case handling. Summary of the Invention

[0005] Embodiments of the present invention provide an insurance customer fraud detection method and system based on artificial intelligence, which can solve the problems in the existing technology.

[0006] In the first aspect of the embodiments of the present invention, There is provided an insurance customer fraud detection method based on artificial intelligence, including: Generating standardized feature data based on insurance application and claim settlement historical data, establishing a non-linear mapping relationship between feature variables and target variables, calculating the partial correlation coefficient between features, identifying the causal relationship between features based on the partial correlation coefficient, and dynamically updating the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; Calculate the temporal intervention effect based on the intermediate nodes of the causal path in the causal graph, and obtain the counterfactual effect through the change in the conditional probability distribution. The temporal intervention effect characterizes the dynamic intervention difference of the feature node on the target node, and the counterfactual effect characterizes the expected change value under path propagation. Identify the triggering factors through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; Calculate the recursive reinforcement transfer coefficient and the temporal dynamic loop strength of each causal path in the causal graph, and calculate the link credibility score; Perform weighted fusion on the abnormal degree of the triggering factor and the link credibility score to generate a warning score; when the warning score exceeds the preset warning threshold, extract the path with the maximum cumulative transfer effect from the causal graph, and construct an evidence chain from the event nodes in the path and their corresponding link credibility scores, and output the fraud risk warning result.

[0007] In an alternative embodiment, Generating a causal graph in the field of insurance fraud includes: Based on the standardized feature data, construct a non-linear mapping relationship, map the feature vector to the fraud risk score space through the basis function, and obtain the initial association between the feature and the target variable; Calculate the partial correlation coefficient between feature pairs based on the initial association, and perform a conditional independence test based on the partial correlation coefficient. The conditional independence test determines the independence score between feature pairs by calculating the negative log-likelihood ratio; Construct an initial causal skeleton graph based on the conditional independence test results, and determine the direction of the causal edge according to the V-structure recognition rule and the directed acyclic constraint to generate an initial causal graph; Dynamically update the initial causal graph using the Bayesian probability framework, and calculate the impact of new data on the causal structure through an incremental scoring mechanism. The incremental scoring mechanism combines the marginal contribution of new data with the historical cumulative score; Optimize the updated causal graph based on the weighted combination of the graph structure complexity and the data fitting degree, and output the final causal graph under the constraints of maintaining causal sufficiency and the Markov property.

[0008] In an alternative embodiment, Calculate the temporal intervention effect based on the intermediate nodes of the causal path, and obtain the counterfactual effect through the change in the conditional probability distribution. The temporal intervention effect characterizes the dynamic intervention difference of the feature node on the target node, and the counterfactual effect characterizes the expected change value under path propagation. Identifying the triggering factors through dynamic weighted fusion and exponential decay accumulation includes: Select node pairs in the causal graph, including feature nodes and target nodes, and extract all possible path sets from the feature nodes to the target nodes; For each path in the path set, identify the set of intermediate nodes in the path. Within a preset time window, calculate the conditional expected value of the target node based on the set of intermediate nodes, and calculate the difference in the intervention effect of the feature node on the target node under different value conditions of the feature node to obtain the time-series intervention effect value; For each path in the path set, obtain the actual observed value of the feature node at the current moment, and set a hypothetical intervention value based on the historical data distribution. Calculate the change in the conditional probability distribution of each intermediate node when the feature node changes from the actual observed value to the hypothetical intervention value according to the propagation structure of the path, and calculate the expected change value of the target node when the value of the feature node changes based on the change in the conditional probability distribution to obtain the counterfactual effect value; Perform weighted fusion on the time-series intervention effect value and the counterfactual effect value, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; perform exponential decay weighting on the fusion effect values at different time points to obtain the time-series cumulative effect value, and the exponential decay weight decreases as the time interval increases; rank the importance of the feature nodes according to the time-series cumulative effect value, and identify the preset number of feature nodes with higher rankings as trigger factors.

[0009] In an alternative embodiment, Performing weighted fusion on the time-series intervention effect value and the counterfactual effect value, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value, including: Set initial weight coefficients for the time-series intervention effect value and the counterfactual effect value respectively, and obtain the structured effect value through weighted summation; extract semantic features from the text description field of the insurance claim record and quantify them as text effect values; Based on the attention mechanism, adaptively fuse the structured effect value and the text effect value, and introduce a feature interaction term to capture the synergistic effect of multi-modal data to obtain the fusion effect value. The adaptive fusion includes: calculating the vector outer product of the structured effect value and the text effect value to obtain a feature interaction matrix, performing a non-linear transformation on the feature interaction matrix to extract interaction features, and constructing a fusion function from the linear combination of the structured effect value, the text effect value and the weighted sum of the interaction features; Construct the confidence interval of the fusion effect value through the Bootstrap resampling method, and set the perturbation upper bound to calculate the maximum change range of the fusion effect value with respect to unobserved confounding factors to obtain the sensitivity index; Construct a reliability evaluation mechanism based on the confidence interval width and sensitivity index, including: normalizing the sensitivity index and confidence interval width and mapping them to the [0,1] interval through an exponential function to obtain a reliability score, and multiplying the reliability score by a time decay factor to dynamically update the weight coefficient of the fusion function; Adopt a time decay mechanism to perform weighted accumulation on the fusion effect values at different times to obtain the final fusion effect value.

[0010] In an alternative implementation, Calculate the recursive reinforcement transfer coefficient and the time-series dynamic loop strength of each causal path in the causal graph, and calculate the link credibility score, including: Calculate the influence value of the nodes in the causal graph, and the influence value is obtained through an exponential decay function based on the in-edge weight, out-edge weight of the node, and the distance from the node to the corresponding edge; Perform recursive transfer reinforcement on the causal path based on the influence value, including adding the transfer coefficient of the previous layer of recursion to the weighted sum of the product of the influence value of all nodes in the current layer and the corresponding node path weight, where the weight of the weighted sum is the decay factor corresponding to the recursion layer; obtain the convergence depth of the recursive transfer reinforcement, and the convergence depth is the recursion layer corresponding to when the difference between the transfer coefficients of adjacent two layers of recursion is less than the preset convergence threshold; Construct a loop time-series state vector, and the loop time-series state vector includes the activity of the nodes in the loop, the edge weight change rate, and the loop integrity; calculate the time-series consistency reinforcement value of the loop based on the loop time-series state vector, and the time-series consistency reinforcement value is obtained through the weighted sum of the time-series difference exponential decay value of the loop time-series state vector and the proportion of the discreteness of each dimension feature of the loop time-series state vector; Calculate the coupling degree between the causal path and the loop, and the coupling degree is determined according to the similarity between the recursive transfer coefficient corresponding to the convergence depth and the time-series consistency reinforcement value of the loop and the overlap degree between the causal path and the loop; Perform adaptive weighted fusion on the recursive transfer coefficient and the time-series consistency reinforcement value of the loop based on the coupling degree to obtain the link credibility score, where the weight of the adaptive weighting is updated by the gradient descent method of the loss function of the historical prediction accuracy.

[0011] In an alternative implementation, The calculation of the recursive reinforcement transfer coefficient includes: Calculate the transfer effect of each recursive layer, including for the node set of each recursive layer, adding the transfer coefficient of the previous layer of recursion to the weighted sum of the product of the influence value of all nodes in the current layer and the corresponding node path weight to obtain the hierarchical transfer strength, where the weight of the weighted sum is the decay factor corresponding to the recursion layer, and accumulating the hierarchical transfer strengths of all recursive layers to obtain the cumulative transfer effect; Optimize the cumulative transmission effect, including calculating the hierarchical transmission intensity ratio of adjacent recursive layers to obtain the inter-layer gain rate, calculating the relative influence distribution entropy of the recursive layer nodes, and determining the optimal recursive depth when the inter-layer gain rate is less than the preset gain threshold and the decay ratio of the distribution entropy is greater than the preset entropy threshold; Perform adaptive adjustment according to the optimal recursive depth, including calculating the relative error of the cumulative transmission effect, determining the depth adjustment amount based on the logarithmic ratio of the relative error to the inter-layer gain rate, adding the depth adjustment amount to the current recursive depth to obtain the updated recursive depth until the difference between the recursive transmission coefficients of adjacent two layers is less than the preset convergence threshold, and obtaining the recursive enhanced transmission coefficient.

[0012] In an alternative embodiment, Extract the path with the maximum cumulative transmission effect from the causal graph, and construct an evidence chain from the event nodes in the path and their corresponding link credibility scores, including: Calculate the cumulative transmission effect of the causal path, including calculating the path transmission intensity according to the weights and distance decays of the edges in the path, and performing time-sequence decay weighted accumulation on the path transmission intensity; Calculate the coefficient of variation of the cumulative transmission effect to obtain the transmission stability score, obtain the weighted transmission effect of the relevant path set to obtain the path correlation enhancement value, and perform weighted fusion on the cumulative transmission effect, the transmission stability score, and the path correlation enhancement value to obtain the comprehensive transmission effect; Perform constraint optimization on the comprehensive transmission effect, including multiplying the link credibility score of the path node by the comprehensive transmission effect to obtain the local optimal score, and performing weighted fusion on the local optimal score and the path coverage to obtain the global score; Fuse the link credibility score and the degree centrality value of the node in the causal path to calculate the node importance weight, the degree centrality value is calculated by weighting the number of incoming edges and the number of outgoing edges of the node, calculate the evidence association strength based on the association degree of adjacent nodes and the node importance weight, and multiply the evidence association strength by the global score to obtain the evidence chain integrity score; Select the path with the maximum evidence chain integrity score as the optimal evidence chain path, and construct an evidence chain from the event nodes in the optimal evidence chain path and their corresponding link credibility scores in the transmission direction order.

[0013] In the second aspect of the embodiments of the present invention, Provide an insurance customer fraud detection system based on artificial intelligence, including: The first unit is used to generate standardized feature data based on insurance application and claim history data, establish a non-linear mapping relationship between feature variables and target variables, calculate the partial correlation coefficient between features, identify the causal relationship between features based on the partial correlation coefficient, and dynamically update the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; The second unit is used to calculate the temporal intervention effect based on the intermediate nodes of the causal path in the causal map, and obtain the counterfactual effect through the change of conditional probability distribution. The temporal intervention effect represents the dynamic intervention difference of the feature node on the target node, and the counterfactual effect represents the expected change value under path propagation. Trigger factors are identified through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; The third unit is used to calculate the recursive reinforcement transfer coefficient and temporal dynamic loop strength of each causal path in the causal map, and calculate the link credibility score; The fourth unit is used to perform weighted fusion on the trigger factors and the abnormality degree of the link credibility score to generate a warning score; when the warning score exceeds a preset warning threshold, extract the path with the maximum cumulative transfer effect from the causal map, construct an evidence chain from the event nodes in the path and their corresponding link credibility scores, and output the fraud risk warning result.

[0014] In the third aspect of the embodiments of the present invention, There is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In the fourth aspect of the embodiments of the present invention, There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] By constructing a causal map in the field of insurance fraud and combining temporal intervention effect and counterfactual effect analysis, the present invention can accurately identify the key trigger factors of fraud risks, improve the accuracy and interpretability of fraud detection. Through the method of dynamic weighted fusion and exponential decay accumulation, the system can effectively capture the characteristics of fraud behavior evolving over time and reduce the false alarm rate.

[0017] The present invention uses a recursive reinforcement transfer coefficient and a time-series dynamic loop strength to evaluate the credibility of causal paths, quantifies the evidence strength of fraud risks through link credibility scoring, making the risk assessment results more objective and reliable. Based on the dynamic update mechanism of Bayesian probability, the identification of causal relationships can be continuously optimized and improved with the accumulation of new data, enhancing the adaptability of the model.

[0018] When generating a warning result, the present invention extracts the path with the maximum cumulative transfer effect as an evidence chain, not only giving the risk warning result, but also clearly showing the complete path and key nodes where the risk occurs, providing detailed evidence support. This way of risk interpretation based on the evidence chain greatly improves the comprehensibility and operability of the risk warning result, helping the insurance company to take more targeted preventive measures. Brief Description of the Drawings

[0019] Figure 1 It is a schematic flowchart of the method for detecting insurance customer fraud based on artificial intelligence according to an embodiment of the present invention; Figure 2 It is a comparison diagram of the time-series cumulative effect scheme of the present invention with static causal inference, regression analysis method, and direct correlation analysis; Figure 3 Analysis diagram of the recursive transfer reinforcement convergence process. Detailed Embodiment

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0022] Figure 1 It is a schematic flowchart of the method for detecting insurance customer fraud based on artificial intelligence according to an embodiment of the present invention, as Figure 1 shown, the method includes: Generating standardized feature data based on insurance application and claim history data, establishing a non-linear mapping relationship between feature variables and target variables, calculating the partial correlation coefficient between features, identifying the causal relationship between features based on the partial correlation coefficient, and dynamically updating the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; Calculate the temporal intervention effect based on the intermediate nodes of the causal path in the causal graph, and obtain the counterfactual effect through the change of conditional probability distribution. The temporal intervention effect characterizes the dynamic intervention difference of the feature node on the target node, and the counterfactual effect characterizes the expected change value under path propagation. Identify the triggering factors through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; Calculate the recursive reinforcement transfer coefficient and temporal dynamic loop strength of each causal path in the causal graph, and calculate the link credibility score; Perform weighted fusion on the triggering factors and the abnormal degree of the link credibility score to generate a warning score; when the warning score exceeds the preset warning threshold, extract the path with the maximum cumulative transfer effect from the causal graph, and construct an evidence chain from the event nodes in the path and their corresponding link credibility scores, and output the fraud risk warning result.

[0023] In an alternative embodiment, generating a causal graph in the field of insurance fraud includes: Based on the standardized feature data, construct a non-linear mapping relationship, map the feature vector to the fraud risk score space through the basis function, and obtain the initial association between the feature and the target variable; Calculate the partial correlation coefficient between feature pairs based on the initial association, and perform a conditional independence test based on the partial correlation coefficient. The conditional independence test determines the independence score between feature pairs by calculating the negative log-likelihood ratio; Construct an initial causal skeleton graph based on the conditional independence test results, and determine the direction of the causal edge according to the V-structure identification rule and the directed acyclic constraint to generate an initial causal graph; Dynamically update the initial causal graph using the Bayesian probability framework, and calculate the impact of new data on the causal structure through an incremental scoring mechanism. The incremental scoring mechanism combines the marginal contribution of new data with the historical cumulative score; Optimize the updated causal graph based on the weighted combination of the graph structure complexity and data fitting degree, and output the final causal graph under the constraints of maintaining causal sufficiency and Markov properties.

[0024] Exemplarily, perform standardized preprocessing on the original insurance data. For continuous features, use the maximum-minimum normalization method to map the values to the interval from zero to one; for categorical features, use one-hot encoding to convert them into numerical representations. Taking auto insurance data as an example, unify features such as vehicle type, accident time, and repair items into standardized feature vectors.

[0025] Construct a non - linear mapping relationship. Select the radial basis function as the basis function and map the standardized feature vectors to the fraud risk score space. In specific implementation, the initial association strength between features and the fraud risk target variable is obtained by calculating the distance between the feature vectors and the pre - determined center points. For example, for auto insurance claim data, representative feature combinations can be selected as the center points based on historical fraud cases.

[0026] Calculate the partial correlation coefficient between feature pairs. For any two features, calculate the degree of correlation between them while controlling the influence of other features. Based on the calculated partial correlation coefficient, judge whether there is a significant correlation between feature pairs by setting a threshold. At the same time, calculate the negative log - likelihood ratio as the statistic for conditional independence test to further verify the independence between feature pairs.

[0027] Construct an initial causal skeleton graph based on the conditional independence test results. Connect undirected edges between significantly correlated feature pairs to form an initial graph structure. Subsequently, according to the V - structure recognition rule, that is, when two uncorrelated nodes have a common child node, determine the direction of the edge. At the same time, impose a directed acyclic constraint to ensure that the final generated graph structure has no cyclic dependencies.

[0028] Adopt the Bayesian probability framework to achieve dynamic update of the causal graph. When new insurance data arrives, calculate the posterior probability of each causal edge and combine it with the historical cumulative probability through weighting. By setting a decay factor, control the retention degree of historical information. For example, for newly discovered fraud patterns, the weight of new data can be appropriately increased.

[0029] Optimize the causal graph. Evaluate the quality of the causal graph from two dimensions: the complexity of the graph structure and the degree of data fitting. The complexity of the structure considers the number of edges and the connectivity of nodes, and the degree of data fitting measures the explanatory ability of the model for the observed data. On the premise of maintaining the constraints of causal sufficiency and Markov property, select the optimal causal structure. Causal sufficiency ensures that all necessary direct causal relationships are included in the graph, and the Markov property ensures that a node only depends on its direct parent nodes.

[0030] Existing insurance fraud detection technologies mainly rely on rule matching and statistical models, such as logistic regression and decision trees, etc. The present invention introduces the radial basis function to construct a non - linear mapping from features to the risk space, enhancing the ability to capture complex fraud patterns. It realizes real - time update of the causal structure through the Bayesian framework, enabling the model to adapt to the emergence of new fraud means. At the same time, based on conditional independence test and V - structure recognition, a causal graph is constructed to deeply reveal the true causal relationships between features, and an evidence chain is constructed to provide an explanatory basis for the early warning results.

[0031] In an alternative embodiment, the temporal intervention effect is calculated based on the intermediate nodes of the causal path, and the counterfactual effect is obtained through the change of the conditional probability distribution. The temporal intervention effect characterizes the dynamic intervention difference of the feature node on the target node, and the counterfactual effect characterizes the expected change value under path propagation. The triggering factors are identified through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value, including: Select node pairs in the causal graph, including feature nodes and target nodes, and extract all possible path sets from the feature nodes to the target nodes; For each path in the path set, identify the set of intermediate nodes in the path. Within a preset time window, calculate the conditional expected value of the target node based on the set of intermediate nodes, and calculate the intervention effect difference of the feature node on the target node under different value conditions of the feature node to obtain the temporal intervention effect value; For each path in the path set, obtain the actual observed value of the feature node at the current moment, and set the assumed intervention value based on the historical data distribution. Calculate the change in the conditional probability distribution of each intermediate node when the feature node changes from the actual observed value to the assumed intervention value according to the propagation structure of the path, and calculate the expected change value of the target node when the value of the feature node changes based on the change in the conditional probability distribution to obtain the counterfactual effect value; Perform weighted fusion on the temporal intervention effect value and the counterfactual effect value, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; perform exponential decay weighting on the fusion effect values at different time points to obtain the temporal cumulative effect value, and the exponential decay weight decreases as the time interval increases; perform importance ranking on the feature nodes according to the temporal cumulative effect value, and identify the preset number of feature nodes with higher rankings as triggering factors.

[0032] Exemplarily, in the causal graph, nodes represent different variables, and edges represent the causal relationships between nodes. For a given pair of feature nodes and target nodes, traverse the causal graph through the depth-first search algorithm to extract all possible paths from the feature nodes to the target nodes. For example, in a certain business scenario, the feature node is the user behavior indicator, the target node is the business conversion indicator, and the intermediate nodes include multiple user status indicators. Suppose three different propagation paths are extracted, and each path contains 2-3 intermediate nodes.

[0033] Within the preset time window, calculate the conditional probability distribution of the intermediate nodes based on historical data. Taking 30 days as the time window, for the intermediate nodes on each path, count the conditional distribution of the intermediate nodes when the feature node takes different value intervals.

[0034] For each path, calculate the difference in the intervention effect of different values of the feature node on the target node. Specifically, by calculating the change in the expected value of the target node when the feature node changes from the low-value interval to the high-value interval, the temporal intervention effect is obtained. At the same time, obtain the actual observed value of the feature node, set a hypothetical intervention value, and calculate the change in the distribution of each intermediate node when the value of the feature node changes, so as to obtain the expected change value of the target node, that is, the counterfactual effect.

[0035] Perform weighted fusion of the temporal intervention effect and the counterfactual effect. The weighting coefficient is determined based on the width of the confidence interval of the effect value and the sensitivity to the perturbation of the feature node. The effect value with a narrower confidence interval and higher sensitivity obtains a greater weight.

[0036] Perform exponential decay weighted accumulation on the fusion effect values at different time points. Set the decay factor to 0.9, then the weight of the effect value t days ago is the t-th power of 0.9. Finally, the cumulative effect value considering time decay is obtained. Sort the feature nodes according to the cumulative effect value, and select the top 10% of the nodes as the key triggering factors.

[0037] The present invention accurately quantifies the causal influence intensity of the feature node on the target node by calculating the weighted fusion of the temporal intervention effect and the counterfactual effect, avoids the bias caused by relying only on a single effect measure, improves the reliability of trigger factor identification; calculates the counterfactual effect based on the change in the conditional probability distribution, fully considers the propagation structure of the causal path, accurately depicts the influence mechanism of the change in the value of the feature node on the target node, and improves the accuracy of effect calculation. Figure 2 This is a comparison chart of the temporal cumulative effect scheme of the present invention with static causal inference, regression analysis method, and direct correlation analysis, as Figure 2 shown. The present invention uses exponential decay weighting (decay factor 0.9) to accumulate historical effects, and the curve shows that it can better capture the long-term effects of sudden events (at T-25 and T-10). The static method cannot correctly reflect the time decay characteristics, resulting in an overestimation of the influence of more distant historical events. The direct correlation method and the regression method cannot effectively capture the long-term continuous influence of sudden events. By using the method of dynamic weighted fusion and exponential decay accumulation, the importance of different types of effects is reasonably balanced, and the time decay factor is considered, making the identification result of the trigger factor more reasonable and having temporal characteristics.

[0038] In an alternative embodiment, performing weighted fusion of the temporal intervention effect value and the counterfactual effect value, and dynamically adjusting the weighting coefficient based on the confidence interval and sensitivity of the effect value includes: Set initial weight coefficients for the temporal intervention effect value and the counterfactual effect value respectively, and obtain the structured effect value through weighted summation; extract semantic features from the text description field of the insurance claim record and quantify them as text effect values; Based on the attention mechanism, adaptively fuse the structured effect value and the text effect value, and introduce a feature interaction term to capture the synergistic effect of multi-modal data to obtain a fused effect value. The adaptive fusion includes: calculating the vector outer product of the structured effect value and the text effect value to obtain a feature interaction matrix, performing a non-linear transformation on the feature interaction matrix to extract interaction features, and constructing a fusion function by combining the linear combination of the structured effect value and the text effect value with the weighted sum of the interaction features; Construct the confidence interval of the fused effect value through the Bootstrap resampling method, and set the perturbation upper bound to calculate the maximum change range of the fused effect value with respect to unobserved confounding factors to obtain a sensitivity index; Construct a reliability evaluation mechanism based on the confidence interval width and the sensitivity index, including: normalizing the sensitivity index and the confidence interval width and mapping them to the [0,1] interval through an exponential function to obtain a reliability score, and multiplying the reliability score by a time decay factor to dynamically update the weight coefficient of the fusion function; Adopt a time decay mechanism to perform weighted accumulation on the fused effect values at different times to obtain the final fused effect value.

[0039] Exemplarily, in insurance claim data analysis, first perform weighted fusion processing on the time-series intervention effect value and the counterfactual effect value. By setting the initial weight coefficient, assign a weight of 0.6 to the time-series intervention effect value and a weight of 0.4 to the counterfactual effect value, and perform weighted summation operation to obtain the structured effect value. At the same time, from the text description of insurance claims records, use a pre-trained language model to extract semantic features, convert the text content into a 300-dimensional vector representation, and map it through a fully connected layer to obtain the text effect value.

[0040] Then, realize the adaptive fusion of the structured effect value and the text effect value. Calculate the vector outer product of the two effect values to generate an interaction feature matrix. Apply the ReLU activation function to this matrix for non-linear transformation to extract effective interaction features. Combine the linear combination of the structured effect value and the text effect value with the weighted sum of the interaction features to construct a fusion function. Specifically, when implementing, the attention mechanism can be used to dynamically adjust the weight coefficients of each part.

[0041] To evaluate the reliability of the fused effect value, use the Bootstrap method to perform 1000 resamplings to calculate the 95% confidence interval. At the same time, by applying a maximum perturbation of 10% to the unobserved confounding factors, calculate the change range of the fused effect value as the sensitivity index. Normalize the confidence interval width by dividing it by the mean value of the effect value to obtain the relative uncertainty. Map the sensitivity index and the relative uncertainty to the 0 to 1 interval through a negative exponential function to obtain the reliability score.

[0042] Based on the reliability assessment results, a time decay mechanism is introduced to weight the historical fusion effect values. The benchmark decay coefficient is set to 0.95, the weight at the current moment is 1, and the weight decays once every time unit going back. Multiply the fusion effect values at each moment by their corresponding weights and sum them up to finally obtain the comprehensive assessment result.

[0043] For example, in a motor vehicle insurance claim case, the sequential intervention effect value is 0.85, the counterfactual effect value is 0.75, and the effect value extracted from the text description is 0.80. After adaptive fusion, the fusion effect value is 0.82, its 95% confidence interval is [0.78, 0.86], and the sensitivity index is 0.08. The reliability score is calculated to be 0.85. Combining with the time decay factor, the fusion effect values at previous moments are weighted and accumulated to finally obtain a stable and reliable assessment result.

[0044] Through the dynamic weight adjustment mechanism, the present invention effectively balances the importance of effect values from different sources and improves the accuracy and reliability of the fusion result. By combining the collaborative analysis of text semantic features and structured data, the complementary information in multi-modal data is fully utilized. The reliability assessment mechanism based on the confidence interval and sensitivity analysis provides a credibility quantification index for the fusion result, which helps to identify and control potential estimation biases. Through adaptive weight update, the system can dynamically adjust the fusion strategy according to the data quality, introduce a time decay mechanism, reasonably balance the influence of historical data and the latest observations, and improve the timeliness and stability of the effect value estimation. The multi-level fusion architecture ensures the robustness of the assessment result and provides a reliable quantitative basis for insurance claim decisions.

[0045] In an alternative embodiment, calculating the recursive reinforcement transfer coefficients and the sequential dynamic loop strength of each causal path in the causal graph, and calculating the link credibility score includes: Calculating the influence value of a node in the causal graph, where the influence value is obtained through an exponential decay function based on the in-edge weight, out-edge weight of the node, and the distance from the node to the corresponding edge; Recursively strengthening the causal path based on the influence value, including adding the transfer coefficient of the previous layer of recursion to the weighted sum of the product of the influence values of all nodes in the current layer and the corresponding node path weights, where the weight of the weighted sum is the decay factor corresponding to the recursion layer; obtaining the convergence depth of the recursive transfer reinforcement, where the convergence depth is the recursion layer corresponding to when the difference between the transfer coefficients of adjacent two layers of recursion is less than a preset convergence threshold; Construct a loop timing state vector, where the loop timing state vector includes the activity of nodes in the loop, the edge weight change rate, and the loop integrity; calculate the timing consistency reinforcement value of the loop based on the loop timing state vector, and the timing consistency reinforcement value is obtained through the weighted sum of the exponential decay value of the timing difference of the loop timing state vector and the proportion of the dispersion degree of each dimension feature of the loop timing state vector; Calculate the coupling degree between the causal path and the loop, where the coupling degree is determined according to the similarity between the recursive transfer coefficient corresponding to the convergence depth and the loop timing consistency reinforcement value and the overlap degree between the causal path and the loop; Based on the coupling degree, perform adaptive weighted fusion on the recursive transfer coefficient and the loop timing consistency reinforcement value to obtain a link credibility score, where the weight of the adaptive weighting is updated by the gradient descent method of the loss function of the historical prediction accuracy.

[0046] Exemplarily, calculate the influence value of nodes in the causal graph. Calculating the influence value of nodes requires considering the in-edge weight, out-edge weight, and the distance from the node to the corresponding edge. Specifically, in implementation, for each node, count the weight values of all its in-edges and out-edges, and set a decay coefficient according to the distance of the edge. The farther the distance, the smaller the decay coefficient. For example, when the distance from node A to edge e1 is 2, the decay coefficient can be set to 0.8, and when the distance is 3, it is set to 0.6. Sum the weights of all edges multiplied by the corresponding decay coefficients to obtain the influence value of the node.

[0047] Perform recursive transfer reinforcement calculation on the causal path. In each layer of recursion, multiply the transfer coefficient of the previous layer by the influence values of all nodes in the current layer, multiply by the corresponding path weight, and then perform weighted summation. The weighting coefficient decays as the recursion layer increases. For example, the weighting coefficient for the first layer of recursion is 0.9, the second layer is 0.81, and the third layer is 0.729. When the difference between the recursive transfer coefficients of adjacent two layers is less than a preset threshold (such as 0.001), record the current recursion layer as the convergence depth.

[0048] Construct a loop timing state vector and calculate the timing consistency reinforcement value of the loop. The loop timing state vector contains three dimensions: node activity, edge weight change rate, and loop integrity. Node activity can be calculated by the interaction frequency of the node within a certain time window; the edge weight change rate reflects the change trend of the edge weight over time; the loop integrity represents the completeness of the nodes and edges in the loop. Calculate the timing difference index based on these three dimension features, and perform weighted summation with the proportion of the dispersion degree of each dimension feature to obtain the timing consistency reinforcement value.

[0049] Calculate the coupling degree of the causal path and the loop. The calculation of the coupling degree needs to consider two aspects: one is the similarity between the recursive transfer coefficient corresponding to the convergence depth and the loop timing consistency strengthening value, and the other is the overlap degree between the causal path and the loop. The similarity can be obtained by calculating the cosine distance between the two values, and the overlap degree is determined by counting the proportion of common nodes and edges.

[0050] Perform adaptive weighted fusion based on the coupling degree. When fusing, weights are assigned to the recursive transfer coefficient and the loop timing consistency strengthening value respectively, and the weight values are updated through gradient descent using the loss function of historical prediction data. For example, when the historical prediction accuracy is high, the weight of the corresponding feature will be increased accordingly. The final weighted fusion result is the link credibility score.

[0051] As Figure 3 As shown in the recursive transfer reinforcement convergence process analysis diagram, it shows the recursive transfer coefficient convergence processes of 5 typical causal paths. The horizontal axis represents the recursive layer (1 - 20), and the vertical axis represents the transfer coefficient value (0.0 - 1.0). Path P1 (influence value 0.872) reaches convergence at the 9th layer (difference 0.0008 < threshold 0.001), and the convergence depth is 9; path P2 (influence value 0.756) converges at the 12th layer, with a depth of 12; path P3 (influence value 0.683) converges at the 7th layer, with a depth of 7; path P4 (influence value 0.921) shows faster convergence and converges at the 5th layer; path P5 (influence value 0.542) has poor convergence and reaches convergence only at the 15th layer. It can be seen from the figure that the recursive layer attenuation coefficient is initially 0.9 and decreases according to 0.9n as the recursive layer increases, and the coefficient difference also decreases until it is less than the preset convergence threshold 0.001. Paths with high influence values (such as P4) usually have faster convergence speeds and higher final transfer coefficients, which proves that the recursive transfer reinforcement algorithm of the present invention can effectively distinguish causal paths of different importance levels. The coefficient values after the final convergence of each path are: P1 (0.8682), P2 (0.7436), P3 (0.6752), P4 (0.9153), P5 (0.5362).

[0052] By introducing the node influence value and the recursive transmission reinforcement mechanism, the present invention can comprehensively capture the deep correlation relationships among the nodes on the causal path, avoiding the limitation of traditional methods that only consider direct connections and ignore indirect influences, and improving the accuracy and integrity of causal relationship analysis. Based on the analysis method of the loop timing state vector, multi-dimensional features such as node activity, edge weight change, and loop integrity are integrated to achieve an accurate characterization of the temporal dynamic characteristics of causal relationships, enhancing the timeliness and dynamic adaptability of the analysis results. An adaptive weighted fusion strategy is adopted, and the weights of each feature are dynamically adjusted in a historical data-driven manner, so that the final link credibility score not only maintains high accuracy but also has good generalization ability and robustness.

[0053] In an alternative embodiment, the calculation of the recursive reinforcement transmission coefficient includes: Calculating the transmission effect of each recursive layer, including for the node set of each recursive layer, adding the weighted sum of the transmission coefficient of the previous recursive layer and the product of the influence value of all nodes in the current layer and the corresponding node path weight, where the weight of the weighted sum is the attenuation factor corresponding to the recursive layer number, and accumulating the hierarchical transmission intensities of all recursive layers to obtain the cumulative transmission effect; Optimizing the cumulative transmission effect, including calculating the ratio of the hierarchical transmission intensities of adjacent recursive layers to obtain the inter-layer gain rate, calculating the relative influence distribution entropy of the recursive layer nodes, and determining the optimal recursive depth when the inter-layer gain rate is less than the preset gain threshold and the attenuation ratio of the distribution entropy is greater than the preset entropy threshold; Performing adaptive adjustment according to the optimal recursive depth, including calculating the relative error of the cumulative transmission effect, determining the depth adjustment amount based on the logarithmic ratio of the relative error to the inter-layer gain rate, adding the depth adjustment amount to the current recursive depth to obtain the updated recursive depth, until the difference between the recursive transmission coefficients of adjacent two layers is less than the preset convergence threshold, to obtain the recursive reinforcement transmission coefficient.

[0054] Exemplarily, the transfer effect of the recursive layer is calculated. In the network structure, starting from the initial node, multiple recursive layers are formed by expanding layer by layer outward. For each recursive layer, it is necessary to obtain the set of all nodes included in this layer. Taking a social network as an example, the initial user is the 0th layer, the users directly associated with it form the 1st layer, and the associated users of the 1st layer (excluding the users that have already appeared) form the 2nd layer, and so on. In practical applications, if the influence value of the initial node is 0.8, the first layer contains 3 nodes with influence values of 0.6, 0.5, and 0.4 respectively, and the corresponding path weights are 0.7, 0.6, and 0.5. Then the calculation process of the transfer intensity of the first layer is as follows: Multiply 0.8 by 0.6, 0.5, and 0.4 respectively, then multiply by the corresponding path weights and sum them up, and finally multiply by the attenuation factor 0.9 of the first layer to obtain the transfer intensity of the first layer as 0.558.

[0055] Calculate the inter-layer gain rate and distribution entropy. Taking the above case as an example, if the transfer intensity of the second layer is 0.223, then the inter-layer gain rate of the first and second layers is 0.223 / 0.558 = 0.4. At the same time, count the influence distribution of the nodes in the second layer and calculate the relative entropy value. Assume that the preset gain threshold is 0.5 and the entropy threshold is 0.6. When the inter-layer gain rate 0.4 is less than 0.5 and the entropy value attenuation ratio 0.7 is greater than 0.6, it can be determined that the optimal recursive depth is 2 layers.

[0056] Perform depth adaptive adjustment. Calculate the relative error of the cumulative transfer effect. If it is 0.15 currently, take the logarithmic ratio of it to the inter-layer gain rate 0.4 to obtain the depth adjustment amount 0.2. The current depth 2 plus the adjustment amount 0.2 gives the updated depth 2.2. Repeat the calculation until the difference between the coefficients of adjacent two layers is less than the preset convergence threshold 0.01 and then stop. Finally, obtain the optimized recursive reinforcement transfer coefficient.

[0057] The present invention accurately depicts the influence transfer process between nodes in the network by calculating the transfer effect layer by layer recursively, avoiding the problem of ignoring the multi-level transfer effect in traditional methods and improving the calculation accuracy of the transfer coefficient; introducing the inter-layer gain rate and distribution entropy to optimize the recursive depth, effectively avoiding the waste of computing resources caused by blindly increasing the number of recursive layers, and improving the computing efficiency while ensuring the accuracy; adopting an adaptive depth adjustment mechanism to dynamically adjust the number of recursive layers according to the convergence situation of the transfer effect, making the finally obtained transfer coefficient more stable and reliable, and enhancing the adaptability and robustness of the algorithm.

[0058] In an alternative embodiment, extracting the path with the maximum cumulative transfer effect from the causal graph and constructing an evidence chain from the event nodes in the path and their corresponding link credibility scores includes: Calculate the cumulative transmission effect of the causal path, including calculating the path transmission intensity according to the weights and distance attenuation of the edges in the path, and performing time-series attenuation weighted accumulation on the path transmission intensity; Calculate the coefficient of variation of the cumulative transmission effect to obtain the transmission stability score, obtain the weighted path correlation enhancement value by weighting the transmission effects of the relevant path sets, and perform weighted fusion on the cumulative transmission effect, transmission stability score, and path correlation enhancement value to obtain the comprehensive transmission effect; Perform constraint optimization on the comprehensive transmission effect, including multiplying the link credibility score of the path node by the comprehensive transmission effect to obtain the local optimal score, and performing weighted fusion on the local optimal score and the path coverage to obtain the global score; Fuse the link credibility score and the degree centrality value of the node in the causal path to calculate the node importance weight. The degree centrality value is weighted and calculated according to the number of incoming edges and outgoing edges of the node. Calculate the evidence association strength based on the association degree of adjacent nodes and the node importance weight, and multiply the evidence association strength by the global score to obtain the evidence chain integrity score; Select the path with the largest evidence chain integrity score as the optimal evidence chain path, and construct the event nodes and their corresponding link credibility scores in the optimal evidence chain path in the order of the transmission direction into an evidence chain.

[0059] Exemplarily, when performing weighted fusion on the abnormal degrees of the triggering factors and the link credibility scores, first determine the abnormal degree through the deviation level from the historical data distribution, that is, collect historical data to construct a normal value distribution interval, set segmented thresholds based on the standard deviation multiple of the current observed value deviating from the distribution median, and map the deviation degree to the 0-1 interval; then perform weighted summation using adaptive weights, and the weight coefficients are dynamically adjusted by calculating the prediction contribution degrees of each indicator to historical fraud cases. When the prediction accuracy of an indicator improves, its weight is increased accordingly, and vice versa; finally, normalize the weighted summation result to obtain the early warning score.

[0060] Calculate the cumulative transmission effect of the paths in the causal graph. Specifically, for each causal path, calculate according to the weight values of the edges in the path. For example, for a path containing nodes A->B->C, the weight of edge AB is 0.8, and the weight of edge BC is 0.7, then the initial transmission intensity is 0.56. At the same time, consider the distance attenuation factor. When the distance between adjacent nodes is 1 unit, the attenuation coefficient is 0.9, when the distance is 2 units, the attenuation coefficient is 0.81, and so on. In addition, for events that occurred earlier in time, introduce time-series attenuation weights. For example, the weight of an event one year ago is 0.8, and the weight of an event two years ago is 0.64, etc. Considering these factors comprehensively to obtain the cumulative transmission effect of the path.

[0061] Calculate the stability of the transmission effect. By analyzing the fluctuations of the path transmission effect in historical data, calculate the coefficient of variation as the stability score. For example, if the standard deviation of the transmission effect of a certain path in the past year is 0.1 and the average value is 0.5, then the coefficient of variation is 0.2. At the same time, count the set of other paths related to the current path, and weight the transmission effect according to the degree of association between paths to obtain the path correlation enhancement value. For example, if the degree of association between path P1 and P2 is 0.6, then the transmission effect of P2 will enhance the effect of P1 with a weight of 0.6. Finally, weight and fuse the cumulative transmission effect, stability score, and correlation enhancement value according to the weights of 0.5, 0.3, and 0.2 to obtain the comprehensive transmission effect.

[0062] Perform constraint optimization. Multiply the link credibility score by the comprehensive transmission effect to obtain the local optimal score. At the same time, consider the coverage of the path, that is, the proportion of key event nodes included in the path. Fuse the local optimal score and the coverage according to the weights of 0.7 and 0.3 to obtain the global score.

[0063] Calculate the node importance. For each node in the path, count the number of incoming edges and outgoing edges. The weight of the incoming edge is 0.4, and the weight of the outgoing edge is 0.6 to obtain the degree centrality value of the node. Fuse the degree centrality value with the link credibility score to obtain the node importance weight. Calculate the evidence association strength based on the degree of association between nodes and the node importance weight. Finally, multiply the evidence association strength by the global score to obtain the evidence chain integrity score.

[0064] Finally, select the path with the highest integrity score as the optimal evidence chain path. For example, there are multiple candidate paths in a causal graph. After the above calculations, the integrity score of path P1 is 0.85, the score of P2 is 0.72, and the score of P3 is 0.68. Then select P1 as the optimal path. Construct the event nodes in P1 and their link credibility scores in the transmission order into the final evidence chain.

[0065] The present invention improves the reliability and accuracy of the evidence chain by introducing a multi-dimensional transmission effect calculation mechanism, comprehensively considering path weights, distance attenuation, and temporal features; the constraint optimization method based on node importance weights and path coverage ensures the integrity and representativeness of the constructed evidence chain, avoiding the omission of important information; the evaluation system using evidence association strength and integrity score realizes the objective quantitative evaluation of the evidence chain, providing a reliable quality guarantee for the construction and application of the evidence chain.

[0066] The present invention also includes: The adaptive determination of the preset time window includes: calculating the state transition probability matrix of the intermediate node set at different time scales, determining the stable period of the node state according to the entropy change rate of the state transition probability; differentiating the conditional expectation sequence of the target node under different time windows, and calculating the relative change rate of the conditional expectation; constructing a time window scoring function based on the product of the stable period of the node state and the relative change rate of the conditional expectation; searching for the optimal time window size through the particle swarm optimization algorithm, where the fitness value of the particle is determined by the time window scoring function, and the position update of the particle is constrained by both the historical optimal solution and the global optimal solution; when the time window scoring gain rate of adjacent iterations is less than the preset threshold, taking the current time window size as the optimal time window.

[0067] The calculation of the reinforcement value includes: constructing a multi-scale feature matrix of the loop state sequence, calculating the short-term fluctuation intensity, the medium-term trend slope, and the long-term cycle characteristics of the activity, the edge weight change rate, and the loop integrity respectively; obtaining the dominant change mode based on the singular value decomposition of the feature matrix, and calculating the contribution degree of each mode; constructing a consistency evaluation matrix according to the temporal correlation of the dominant change mode, and calculating the matrix elements through the mutual information and phase synchronization measure between the modes; performing spectral clustering on the consistency evaluation matrix to extract a subset of strongly correlated modes; calculating a comprehensive consistency index based on the subset of strongly correlated modes, which takes into account both the stability within the mode and the synergy between the modes; taking the weighted product of the comprehensive consistency index and the loop integrity as the improved temporal consistency reinforcement value.

[0068] In the second aspect of the embodiments of the present invention, an insurance customer fraud detection system based on artificial intelligence is provided, and the system includes: The first unit is used to generate standardized feature data based on the insurance application and claim history data, establish a non-linear mapping relationship between the feature variables and the target variable, calculate the partial correlation coefficient between the features, identify the causal relationship between the features based on the partial correlation coefficient, and dynamically update the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; The second unit is used to calculate the temporal intervention effect based on the intermediate nodes of the causal path in the causal map, and obtain the counterfactual effect through the change of the conditional probability distribution. The temporal intervention effect characterizes the dynamic intervention difference of the feature node on the target node, and the counterfactual effect characterizes the expected change value under the path propagation. The triggering factor is identified through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; The third unit is used to calculate the recursive reinforcement transfer coefficient and the temporal dynamic loop strength of each causal path in the causal map, and calculate the link credibility score; The fourth unit is configured to perform weighted fusion on the triggering factor and the degree of abnormality of the link credibility score to generate a warning score; when the warning score exceeds a preset warning threshold, extract the path with the maximum cumulative transmission effect from the causal graph, construct an evidence chain from the event nodes in the path and their corresponding link credibility scores, and output a fraud risk warning result.

[0069] In the third aspect of the embodiments of the present invention, A kind of electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0070] In the fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0071] The present invention may be a method, apparatus, system and / or computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for performing various aspects of the present invention are uploaded.

[0072] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An insurance customer fraud detection method based on artificial intelligence, characterized in that: include: Generate standardized feature data based on insurance application and claims historical data, Establish a nonlinear mapping relationship between feature variables and target variables, calculate the partial correlation coefficient between features, identify the causal relationship between features based on the partial correlation coefficient, and dynamically update the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; In the causal graph, the temporal intervention effect is calculated based on the propagation of the intermediate nodes of the causal path, and the counterfactual effect is obtained through the change of the conditional probability distribution. The temporal intervention effect represents the dynamic intervention difference of the characteristic node on the target node, and the counterfactual effect represents the expected change value under the path propagation. The triggering factor is identified through dynamic weighted fusion and exponential decay accumulation, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; Calculating the recursive reinforcement transfer coefficient and the temporal dynamic loop strength of each causal path in the causal graph, and calculating the link credibility score; The trigger factor is weighted and integrated with the abnormal degree of the link credibility score to generate an early warning score; When the warning score exceeds the preset warning threshold, the path with the maximum cumulative transmission effect is extracted from the causal graph, the event nodes in the path and their corresponding link credibility scores are constructed into an evidence chain, and the fraud risk warning result is output.

2. The method according to claim 1, characterized in that: Generating a causal graph in the insurance fraud domain involves: Based on the standardized feature data, a nonlinear mapping relationship is constructed, and the feature vector is mapped to the fraud risk score space through a basis function to obtain an initial association between the feature and the target variable; Calculating a partial correlation coefficient between feature pairs based on the initial correlation, and performing a conditional independence test based on the partial correlation coefficient, wherein the conditional independence test determines an independence score between feature pairs by calculating a negative log-likelihood ratio; Based on the conditional independence test results, an initial causal skeleton graph is constructed, and the direction of the causal edge is determined according to the V structure recognition rule and directed acyclic constraints to generate an initial causal graph; The initial causal graph is dynamically updated using a Bayesian probability framework, and the impact of the new data on the causal structure is calculated through an incremental scoring mechanism that combines the marginal contribution of the new data with the historical cumulative score; The updated causal graph is optimized based on a weighted combination of graph structure complexity and data fitting degree, and the final causal graph is output while maintaining the constraints of causal sufficiency and Markov properties.

3. The method according to claim 1, characterized in that: The temporal intervention effect is calculated based on the propagation of the intermediate nodes of the causal path, and the counterfactual effect is obtained through the change of the conditional probability distribution. The temporal intervention effect represents the dynamic intervention difference of the characteristic node on the target node. The counterfactual effect represents the expected change value under the path propagation. The triggering factors are identified through dynamic weighted fusion and exponential decay accumulation. The weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value, including: Selecting a node pair in the causal graph, including a feature node and a target node, and extracting a set of all possible paths from the feature node to the target node; For each path in the path set, identify the set of intermediate nodes in the path, calculate the conditional expected value of the target node based on the set of intermediate nodes within a preset time window, and calculate the difference in intervention effect of the characteristic node on the target node under different value conditions to obtain the temporal intervention effect value; For each path in the path set, the actual observed value of the feature node at the current moment is obtained, and the hypothetical intervention value is set based on the historical data distribution. The conditional probability distribution change of each intermediate node when the feature node changes from the actual observed value to the hypothetical intervention value is calculated according to the propagation structure of the path. The expected change value of the target node when the feature node value changes is calculated based on the conditional probability distribution change to obtain the counterfactual effect value; The temporal intervention effect value and the counterfactual effect value are weightedly fused, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; the fused effect values ​​at different time points are exponentially decay weighted to obtain a temporal cumulative effect value, and the exponential decay weight decreases as the time interval increases; the feature nodes are ranked in importance according to the temporal cumulative effect value, and a preset number of feature nodes with the highest ranking are identified as trigger factors.

4. The method according to claim 3, characterized in that The temporal intervention effect value and the counterfactual effect value are weighted and merged, and the weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value, including: Initial weight coefficients are set for the temporal intervention effect value and the counterfactual effect value respectively, and a structured effect value is obtained by weighted summation; semantic features are extracted from the text description field of the insurance claim record and quantified into a text effect value; Based on the attention mechanism, the structured effect value and the text effect value are adaptively fused, and the feature interaction term is introduced to capture the synergistic effect of multimodal data to obtain the fused effect value. The adaptive fusion includes: calculating the vector outer product of the structured effect value and the text effect value to obtain a feature interaction matrix, performing nonlinear transformation on the feature interaction matrix to extract interaction features, and constructing a fusion function by combining the linear combination of the structured effect value and the text effect value with the weighted sum of the interaction features; The confidence interval of the fusion effect value is constructed by the Bootstrap resampling method, and the upper limit of the disturbance is set to calculate the maximum change amplitude of the fusion effect value to the unobserved confounding factors to obtain the sensitivity index; A reliability evaluation mechanism is constructed based on the confidence interval width and the sensitivity index, including: normalizing the sensitivity index and the confidence interval width and mapping them to the [0,1] interval through an exponential function to obtain a reliability score, and dynamically updating the weight coefficient of the fusion function by multiplying the reliability score with the time decay factor; The time decay mechanism is used to perform weighted accumulation on the fusion effect values ​​at different times to obtain the final fusion effect value.

5. The method according to claim 1, characterized in that Calculating the recursive reinforcement transfer coefficient and the temporal dynamic loop strength of each causal path in the causal graph, and calculating the link credibility score includes: Calculate the influence value of the node in the causal graph, where the influence value is obtained through an exponential decay function according to the node's incoming edge weight, outgoing edge weight, and the distance from the node to the corresponding edge; Recursively strengthening the causal path based on the influence value includes adding the weighted sum of the product of the transfer coefficient of the previous layer recursively and the influence values ​​of all nodes in the current layer multiplied by the corresponding node path weight, wherein the weight of the weighted sum is the attenuation factor corresponding to the number of recursive layers; obtaining the convergence depth of the recursive transfer strengthening, wherein the convergence depth is the number of recursive layers corresponding to when the difference between the recursive transfer coefficients of two adjacent layers is less than a preset convergence threshold; Constructing a loop timing state vector, wherein the loop timing state vector includes the activity of nodes in the loop, the rate of change of edge weights, and the integrity of the loop; calculating a timing consistency reinforcement value of the loop based on the loop timing state vector, wherein the timing consistency reinforcement value is obtained by weighted summing the timing difference exponential decay value of the loop timing state vector and the discrete degree proportion of each dimensional feature of the loop timing state vector; Calculating the coupling degree between the causal path and the loop, wherein the coupling degree is determined according to the similarity between the recursive transfer coefficient corresponding to the convergence depth and the loop temporal consistency reinforcement value and the overlap between the causal path and the loop; Based on the coupling degree, the recursive transfer coefficient and the loop timing consistency enhancement value are adaptively weighted and fused to obtain a link credibility score, wherein the weight of the adaptive weighting is updated by a gradient descent method of a loss function of historical prediction accuracy.

6. The method according to claim 5, characterized in that The calculation of the recursive reinforcement transfer coefficient includes: Calculate the transfer effect of each recursive layer, including for each recursive layer node set, add the weighted sum of the transfer coefficient of the previous recursive layer and the influence value of all nodes in the current layer multiplied by the corresponding node path weight product to obtain the hierarchical transfer strength, where the weight of the weighted sum is the attenuation factor corresponding to the number of recursive layers, and accumulate the hierarchical transfer strength of all recursive layers to obtain the cumulative transfer effect; Optimizing the cumulative transfer effect, including calculating the inter-layer gain rate by calculating the ratio of the layer transfer strengths of adjacent recursive layers, calculating the relative influence distribution entropy of the recursive layer nodes, and determining the optimal recursive depth when the inter-layer gain rate is less than a preset gain threshold and the attenuation ratio of the distribution entropy is greater than a preset entropy threshold; Adaptive adjustment is performed according to the optimal recursive depth, including calculating the relative error of the cumulative transfer effect, determining the depth adjustment amount based on the logarithmic ratio of the relative error to the inter-layer gain rate, adding the depth adjustment amount to the current recursive depth to obtain an updated recursive depth, until the difference between the recursive transfer coefficients of two adjacent layers is less than the preset convergence threshold, and the recursive enhancement transfer coefficient is obtained.

7. The method according to claim 1, characterized in that Extracting the path with the maximum cumulative transmission effect from the causal graph, and constructing the event nodes in the path and their corresponding link credibility scores into an evidence chain includes: Calculating the cumulative transfer effect of the causal path, including calculating the path transfer strength according to the weights of the edges in the path and the distance attenuation, and performing time-series attenuation weighted accumulation on the path transfer strength; Calculating the coefficient of variation of the cumulative transmission effect to obtain a transmission stability score, obtaining the weighted transmission effect of the relevant path set to obtain a path correlation enhancement value, and weightedly fusing the cumulative transmission effect, the transmission stability score and the path correlation enhancement value to obtain a comprehensive transmission effect; Performing constrained optimization on the comprehensive transfer effect, including multiplying the link credibility score of the path node by the comprehensive transfer effect to obtain a local optimal score, and weightedly fusing the local optimal score with the path coverage to obtain a global score; The node importance weight is calculated by integrating the link credibility score and the degree centrality value of the node in the causal path. The degree centrality value is weighted according to the number of incoming edges and the number of outgoing edges of the node. The evidence association strength is calculated based on the association degree of the adjacent nodes and the node importance weight. The evidence association strength is multiplied by the global score to obtain the evidence chain integrity score; The path with the largest evidence chain integrity score is selected as the optimal evidence chain path, and the event nodes in the optimal evidence chain path and their corresponding link credibility scores are constructed into an evidence chain in the order of transmission direction.

8. An insurance customer fraud detection system based on artificial intelligence, used to implement the method of any one of claims 1 to 7, characterized in that: include: The first unit is used to generate standardized feature data based on insurance application and claim history data. Establish a nonlinear mapping relationship between feature variables and target variables, calculate the partial correlation coefficient between features, identify the causal relationship between features based on the partial correlation coefficient, and dynamically update the causal relationship through Bayesian probability calculation to generate a causal map in the field of insurance fraud; The second unit is used to calculate the temporal intervention effect based on the propagation of the intermediate nodes of the causal path in the causal graph, and obtain the counterfactual effect through the change of the conditional probability distribution. The temporal intervention effect represents the dynamic intervention difference of the characteristic node on the target node. The counterfactual effect represents the expected change value under the path propagation. The triggering factor is identified through dynamic weighted fusion and exponential decay accumulation. The weighting coefficient is dynamically adjusted based on the confidence interval and sensitivity of the effect value; The third unit is used to calculate the recursive reinforcement transfer coefficient and the temporal dynamic loop strength of each causal path in the causal graph, and calculate the link credibility score; The fourth unit is used to perform weighted fusion of the trigger factor and the abnormal degree of the link credibility score to generate a warning score; When the warning score exceeds the preset warning threshold, the path with the maximum cumulative transmission effect is extracted from the causal graph, the event nodes in the path and their corresponding link credibility scores are constructed into an evidence chain, and the fraud risk warning result is output.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images