Insurance fraud prevention and detection system based on AI

Abstract

The invention relates to the technical field of insurance fraud identification, in particular to an AI-based insurance fraud prevention and detection system. Comprising a data collection module, a data preprocessing module, a model training module and an interpretability analysis module, the data preprocessing module is connected with the data collection module, and the model training module is connected with the data preprocessing module and the interpretability analysis module; the data collection module is used for acquiring data from various data sources, processing the data and then integrating the data; the data preprocessing module is used for preprocessing the integrated data to obtain normalized data; the model training module is used for constructing a fraud detection model and making a decision for the preprocessed data; the interpretability analysis module is used for displaying the decision result and interacting with the user; through the mode, the decision-making process is provided after fraud detection is carried out on the data, and an insurance company and a supervision mechanism can better understand and trust a detection result.

Description

Technical Field

[0001] The present invention relates to the technical field of insurance fraud identification, and in particular to an insurance fraud prevention and detection system based on AI. Background Art

[0002] Insurance fraud is a long-standing problem faced by the insurance industry, which has brought huge economic burdens to insurance companies and consumers. Fraudulent behaviors include false reporting of losses, falsification of accidents, exaggeration of claim amounts, falsification of victim identities, etc., which not only increase the operating costs of insurance companies, but also lead to higher premiums for legitimate policyholders, affecting the healthy development of the entire insurance market. Traditional insurance fraud detection methods mainly rely on rule engines and expert experience. A rule engine is a system that identifies suspicious behaviors based on predefined rules and conditions, and is usually designed by an expert team of an insurance company based on historical experience and known fraud patterns. These rules may include abnormally large claims, frequent claims, high-risk behaviors in a specific geographic area, etc. However, this rule-based system has obvious limitations. First, rule engines are usually static, and once set up, they are difficult to adapt to changing fraud methods and emerging fraud patterns. Fraudsters can analyze and understand these rules to design evasion strategies, making the rule engine ineffective. Second, rule engines require constant manual intervention and updates to keep up with the evolution of fraud methods, which not only increases operational complexity but also leads to reduced detection efficiency. Additionally, rule engines often fall short when it comes to handling complex, cross-domain frauds, which may involve multiple steps and different types of data, which is beyond the detection capabilities of simple rules.

[0003] The development of artificial intelligence (AI) technology, especially the application of machine learning and deep learning, has brought new solutions to insurance fraud detection. AI-based fraud detection systems can automatically identify complex patterns and abnormal behaviors by deeply analyzing large-scale historical data. Machine learning algorithms, especially supervised learning and unsupervised learning, can train models based on a large number of annotated fraud cases to identify new fraud behaviors. In addition, AI systems can continuously optimize themselves and adapt to new fraud methods as data increases and time passes. However, although AI-based fraud detection systems have shown great potential, they also face some challenges: the characteristics of AI models make their decision-making process difficult to explain and can only provide partial explanations. When faced with these complex models, insurance companies and regulators may question their results, thereby limiting the scope of application of AI models.

[0004] Therefore, it is very necessary to propose an AI-based insurance fraud prevention and detection system that provides a decision-making process after fraud detection on data, so that insurance companies and regulators can better understand and trust the detection results. Summary of the invention

[0005] The purpose of the present invention is to provide an AI-based insurance fraud prevention and detection system, which solves the problem that the prior art can only provide partial explanations after fraud detection on data, and insurance companies and regulatory agencies may question the results, thus limiting the scope of application of AI models.

[0006] To achieve the above-mentioned purpose, the present invention adopts an AI-based insurance fraud prevention and detection system, comprising a data collection module, a data preprocessing module, a model training module and an interpretability analysis module, wherein the data preprocessing module is connected to the data collection module, and the model training module is connected to the data preprocessing module and the interpretability analysis module respectively; The data collection module is used to obtain data from various data sources and integrate them after processing; The data preprocessing module is used to preprocess the integrated data to obtain normalized data; The model training module is used to build a fraud detection model and make decisions based on the preprocessed data; The explainability analysis module is used to display decision results and interact with users.

[0007] Among them, the data collection module includes a multi-source data acquisition unit, a preprocessing unit, a feature extraction unit, a data storage unit, a data protection unit and a real-time data processing unit. The preprocessing unit is respectively connected to the multi-source data acquisition unit and the feature extraction unit, the data storage unit is connected to the feature extraction unit, and the data protection unit is respectively connected to the data storage unit and the real-time data processing unit.

[0008] Wherein, the multi-source data acquisition unit is used to collect data from multiple heterogeneous data sources; The pre-processing unit is used to clean, convert and enhance the collected data; The feature extraction unit is used to extract features from complex relational data and perform data integration; The data storage unit is used to store the integrated data; The data protection unit is used to encrypt the transmission data; The real-time data processing unit is used to integrate the data processing framework and process data in real time.

[0009] The data preprocessing module includes a data cleaning unit, a feature selection unit and a data segmentation unit. The data cleaning unit is connected to the data collection module, and the feature selection unit is connected to the data cleaning unit and the data segmentation unit respectively.

[0010] Wherein, the data cleaning unit is used to remove noise and outliers from the data; The feature selection unit is used to select features to obtain a data set, and to expand and transform the data set to obtain a standardized data set; The data segmentation unit is used to segment the data set into a training set, a validation set and a test set.

[0011] Among them, the model training module includes a deep learning model training unit, a graph neural network model, a model updating unit, an optimization unit and a model integration unit. The deep learning model training unit is connected to the data preprocessing module, the graph neural network model is respectively connected to the deep learning model training unit and the model updating unit, and the optimization unit is respectively connected to the model updating unit and the model integration unit.

[0012] The deep learning model training unit is used to provide a deep learning model and use the acquired training set, validation set and test set to perform the first training on the fraud detection model; The graph neural network model is used to provide a graph neural network model, modeling through nodes and edges, capturing high-order relationship features in the data, performing secondary training on the fraud detection model, and making decisions based on the data; The model updating unit is used to collect update data, annotate the update data, and periodically retrain the fraud detection model using the update data; The optimization unit is used to optimize the parameter configuration of the fraud detection model; The model integration unit is used to perform multi-model integration on the fraud detection model.

[0013] Among them, the explainability analysis module includes an explainability unit, a decision visualization unit, a text description generation unit, an interaction unit and a feedback unit. The explainability unit is connected to the model training module, the decision visualization unit is connected to the explainability unit and the text description generation unit respectively, the interaction unit is connected to the text description generation unit, and the feedback unit is connected to the interaction unit.

[0014] Wherein, the explainability unit is used to output the decision basis of the fraud detection model; The decision visualization unit is used to display the working mechanism of the fraud detection model in the form of a visualization chart and provide a basis for decision-making; The text description generating unit is used to generate a text description to explain the decision of the fraud detection model; The interactive unit is used to provide an interactive interface, record parameter settings selected by different users, and conduct comparative analysis to view decision differences between different settings; The feedback unit is used to collect user feedback data on the fraud detection model.

[0015] An AI-based insurance fraud prevention and detection system of the present invention adopts the data collection module, the data preprocessing module, the model training module and the explainability analysis module to perform insurance fraud prevention and detection operations, wherein: the data collection module is used to obtain data from multiple data sources and integrate them after processing; the data preprocessing module is used to preprocess the integrated data to obtain normalized data; the model training module is used to build a fraud detection model and make decisions based on the preprocessed data; the explainability analysis module is used to display the decision results and interact with the user; the decision-making process is provided after fraud detection of the data, so that insurance companies and regulators can better understand and trust the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 It is a schematic diagram of the principle of the AI-based insurance fraud prevention and detection system of the present invention.

[0018] Figure 2 It is a schematic diagram of the principle of the data collection module of the present invention.

[0019] Figure 3 It is a schematic diagram of the principle of the data preprocessing module of the present invention.

[0020] Figure 4 It is a schematic diagram of the principle of the model training module of the present invention.

[0021] Figure 5 It is a schematic diagram of the principle of the explainability analysis module of the present invention.

[0022] Figure 6 It is a schematic diagram of the principle of the report generation and feedback module of the present invention.

[0023] 100-data collection module, 101-multi-source data acquisition unit, 102-preprocessing unit, 103-feature extraction unit, 104-data storage unit, 105-data protection unit, 106-real-time data processing unit, 200-data preprocessing module, 201-data cleaning unit, 202-feature selection unit, 203-data segmentation unit, 300-model training module, 301-deep learning model training unit, 302-graph neural network model, 303-model update unit, 304-optimization unit, 305-model integration unit, 400-interpretability analysis module, 401-interpretability unit, 402-decision visualization unit, 403-text description generation unit, 404-interaction unit, 500-report generation and feedback module, 501-fraud detection report generation unit, 502-dynamic report update unit, 503-report output unit, 504-audit unit, 505-model improvement unit, 506-user experience optimization unit. DETAILED DESCRIPTION

[0024] See also Figure 1 to Figure 6 ,in Figure 1 This is a schematic diagram of the principle of the AI-based insurance fraud prevention and detection system. Figure 2 This is a schematic diagram of the data collection module. Figure 3 This is a schematic diagram of the principle of the data preprocessing module. Figure 4 This is a schematic diagram of the model training module. Figure 5 It is a schematic diagram of the principle of the explainability analysis module. Figure 6 It is a schematic diagram of the principle of the report generation and feedback module.

[0025] The present invention provides an insurance fraud prevention and detection system based on AI, comprising a data collection module 100, a data preprocessing module 200, a model training module 300 and an interpretability analysis module 400, wherein the data preprocessing module 200 is connected to the data collection module 100, and the model training module 300 is connected to the data preprocessing module 200 and the interpretability analysis module 400 respectively; The data collection module 100 is used to obtain data from various data sources and integrate them after processing; The data preprocessing module 200 is used to preprocess the integrated data to obtain normalized data; The model training module 300 is used to build a fraud detection model and make decisions based on the preprocessed data; The explainability analysis module 400 is used to display decision results and interact with users.

[0026] In this embodiment, the data collection module 100, the data preprocessing module 200, the model training module 300 and the explainability analysis module 400 are used to perform insurance fraud prevention and detection operations, wherein: the data collection module 100 obtains data from multiple data sources and integrates them after processing; the data preprocessing module 200 preprocesses the integrated data to obtain normalized data; the model training module 300 builds a fraud detection model and makes decisions based on the preprocessed data; the explainability analysis module 400 displays the decision results and interacts with the user; the decision-making process is provided after fraud detection of the data, so that insurance companies and regulators can better understand and trust the detection results.

[0027] Furthermore, the data collection module 100 includes a multi-source data acquisition unit 101, a preprocessing unit 102, a feature extraction unit 103, a data storage unit 104, a data protection unit 105 and a real-time data processing unit 106, the preprocessing unit 102 is respectively connected to the multi-source data acquisition unit 101 and the feature extraction unit 103, the data storage unit 104 is connected to the feature extraction unit 103, and the data protection unit 105 is respectively connected to the data storage unit 104 and the real-time data processing unit 106.

[0028] Furthermore, the multi-source data acquisition unit 101 is used to collect data from multiple heterogeneous data sources; The pre-processing unit 102 is used to clean, convert and enhance the collected data; The feature extraction unit 103 is used to extract features from complex relational data and perform data integration; The data storage unit 104 is used to store the integrated data; The data protection unit 105 is used to encrypt the transmission data; The real-time data processing unit 106 is used to integrate the data processing framework and process data in real time.

[0029] In this embodiment, the multi-source data collection unit 101 collects data from multiple heterogeneous data sources, including structured data: such as customer information, claim records, financial data, etc.; unstructured data: such as social media content, text records, audio and video files, etc.; semi-structured data: such as log files in JSON and XML formats; blockchain data: through the blockchain network interface, obtains tamper-proof insurance transaction records. The pre-processing unit 102 cleans, converts, and enhances the collected data. After data collection, the module uses ETL technology to perform the following processing on the data: data cleaning: deduplication, error correction, and missing value filling; data conversion: standardization and normalization of data, processing data from different sources in a unified format; data enhancement: using natural language processing technology to perform semantic analysis and extract key features in the text. The feature extraction unit 103 extracts features from complex relational data and integrates data. In order to reduce data dimensions and retain important information, the following technologies are used: principal component analysis: used to reduce data dimensions and remove redundant features; autoencoder: deep learning model, automatically extracts and compresses important features; graph neural network (GNN): extracts features from complex relational data, such as the network relationship between customers and agents. The data storage unit 104 stores the integrated data, using a combination of graph database and relational database for data storage: graph database: used to store and query complex relational data, such as social networks and transaction networks; relational database: used to store structured data, such as customer information and claim records. The data protection unit 105 encrypts data for transmission. To ensure data security and privacy, the module implements the following measures: encrypted transmission: using TLS / SSL protocol to ensure data security during transmission; storage encryption: using AES encryption technology to protect static data; access control: role-based and attribute-based access control mechanism to ensure that only authorized users can access sensitive data. The real-time data processing unit 106 integrates a data processing framework to process data in real time and implements the following functions: Low-latency processing: real-time processing and analysis of data streams to meet the needs of instant detection of insurance fraud; Event-driven: automatic triggering of alarms or further analysis based on abnormal patterns detected in real time. The data collection module 100 is responsible for acquiring, cleaning, converting and integrating data from multiple data sources to ensure the efficiency and accuracy of subsequent model processing.

[0030] Furthermore, the data preprocessing module 200 includes a data cleaning unit 201, a feature selection unit 202 and a data segmentation unit 203, the data cleaning unit 201 is connected to the data collection module 100, and the feature selection unit 202 is connected to the data cleaning unit 201 and the data segmentation unit 203 respectively.

[0031] Furthermore, the data cleaning unit 201 is used to remove noise and outliers from the data; The feature selection unit 202 is used to select features, obtain a data set, and expand and transform the data set to obtain a standardized data set; The data segmentation unit 203 is used to segment the data set into a training set, a validation set and a test set.

[0032] In this embodiment, the data cleaning unit 201 removes noise and outliers from the data, and the main operations include: outlier detection and processing: using Z-score and IQR methods to detect outliers in numerical data, and marking or removing abnormal data that exceeds the threshold based on statistical rules; missing value processing: for missing values ​​in the data set, a variety of methods are used to fill them, such as mean filling, KNN interpolation or prediction filling based on regression models; data deduplication: efficient identification and removal of duplicate records through algorithms such as hash functions or Bloom filters to reduce data redundancy; format normalization: unify the formats of data from different sources, such as date format, currency unit, etc., to ensure the consistency of data in subsequent processing. Feature selection is a key step in optimizing model performance. It reduces the dimension of data and improves the generalization ability of the model by screening out the most representative features. The main technologies include: filtering method: screening out features with high correlation by calculating the correlation between features and target variables; wrapping method: using algorithms such as recursive feature elimination to gradually eliminate unimportant features by training the model and evaluating its performance; embedding method: automatically selecting features during model training, such as L1 regularization and feature importance ranking in decision tree models; feature importance scoring: using an integrated learning model to calculate the importance score of each feature, which is used as the basis for feature selection.

[0033] Formula example: Assume that the feature matrix in the data set is , the target variable is , each feature can be calculated through correlation analysis With the target variable The correlation coefficient , the formula is as follows:

[0034] According to the correlation coefficient The most important features are selected for model training.

[0035] Data enhancement is an important step to improve the robustness and generalization ability of the model by expanding and transforming the training data set. The main technologies include: oversampling and undersampling: for class-imbalanced data sets, use SMOTE to oversample minority class samples, or balance the class distribution of the data set by random undersampling; data generation model: use generative adversarial networks to generate virtual data similar to the original data to expand the diversity of the training data set; noise injection: randomly add tiny noise to the feature value to enhance the robustness of the model to data fluctuations; data transformation: for specific types of data, perform data transformations such as rotation, scaling, and translation, which is particularly suitable for processing images or time series data. The feature selection unit 202 selects features to obtain a data set, expands and transforms the data set, and obtains a standardized data set. The data segmentation unit 203 segments the data set into a training set, a validation set, and a test set, and uses a cross-validation method, such as K-fold cross-validation, to ensure the robustness of the model. The ratio of data segmentation can generally be 70% training set, 15% validation set, and 15% test set.

[0036] Through the above data preprocessing steps, the data preprocessing module 200 ensures the quality of the data input into the fraud detection model, thereby improving the training effect and prediction performance of the fraud detection model. In this process, not only the noise and outliers in the data are eliminated, but also the generalization ability of the fraud detection model is improved through feature selection and data enhancement technology, so that when facing complex fraudulent behaviors, it can still maintain efficient and accurate detection capabilities.

[0037] Furthermore, the model training module 300 includes a deep learning model training unit 301, a graph neural network model 302, a model updating unit 303, an optimization unit 304 and a model integration unit 305, the deep learning model training unit 301 is connected to the data preprocessing module 200, the graph neural network model 302 is respectively connected to the deep learning model training unit 301 and the model updating unit 303, and the optimization unit 304 is respectively connected to the model updating unit 303 and the model integration unit 305.

[0038] Furthermore, the deep learning model training unit 301 is used to provide a deep learning model and use the acquired training set, validation set and test set to perform initial training on the fraud detection model; The graph neural network model 302 is used to provide a graph neural network model 302, modeling through nodes and edges, capturing high-order relationship features in the data, performing secondary training on the fraud detection model, and making decisions based on the data; The model updating unit 303 is used to collect update data, annotate the update data, and use the update data to periodically retrain the fraud detection model; The optimization unit 304 is used to optimize the parameter configuration of the fraud detection model; The model integration unit 305 is used to perform multi-model integration on the fraud detection model.

[0039] In this embodiment, a large-scale annotated data set is trained and optimized through deep learning and graph neural network, so as to construct an efficient fraud detection model. The deep learning model training unit 301 provides a deep learning model, and uses the obtained training set, validation set and test set to train the fraud detection model for the first time. Several mainstream deep learning models can be used for training: Multilayer Perceptron: used to process structured data, extract nonlinear features in the data through multiple fully connected layers; Convolutional Neural Network: mainly used for fraud detection of images or similar two-dimensional data, such as analysis of image evidence or document scans; Recurrent Neural Network and its improved version Long Short-Term Memory Network and Gate Recurrent Unit: used to process time series data, such as analysis of customer transaction history; The training process uses stochastic gradient descent and its optimized version Adam optimizer, and gradually adjusts the model parameters under the guidance of the loss function to minimize the prediction error of the model for the training data. The choice of loss function depends on the specific task type, such as the cross entropy loss function commonly used in classification tasks, and the mean square error is used in regression tasks. The graph neural network model 302 provides a graph neural network model 302, which is modeled by nodes and edges, captures high-order relationship features in the data, performs secondary training on the fraud detection model, and makes decisions based on the data, using the following architecture: Graph convolutional network: uses convolution operations to propagate information in the graph structure, which is suitable for processing static graph data; Graph attention network: weighted averages adjacent nodes through the self-attention mechanism, so as to more accurately capture the relationship between nodes; Dynamic graph neural network: processes time-varying graph data, and can model the evolution of fraud behavior in the time dimension. The model update unit 303 collects updated data, annotates the updated data, and uses the updated data to periodically retrain the fraud detection model. The retraining process is as follows: Data collection and label update: regularly collect the latest insurance data, and annotate the new data through manual or semi-automatic methods; Model retraining: use the updated data set to incrementally train the model, while retaining the learning results of the old data, introduce the features of the new data; Performance evaluation and deployment: use the validation set and the test set to evaluate the performance of the retrained model to ensure that its performance on the new data is better than or not lower than the old model, and then deploy it to the production environment. The optimization unit 304 optimizes the parameter configuration of the fraud detection model and introduces the AutoML technology to automatically optimize the hyperparameter settings of the model. AutoML is mainly responsible for the following tasks: Hyperparameter search: Find the optimal combination in the model hyperparameter space through grid search, random search, and more advanced Bayesian optimization methods; Model architecture search: For deep learning models, the AutoML system can automatically search and adjust the number of network layers, the number of neurons in each layer, the activation function and other architecture parameters to find the network structure with the best performance; Automatic feature engineering: AutoML can also automatically generate and filter features to further improve the prediction ability of the model.The model integration unit 305 performs multi-model integration on the fraud detection model and supports multi-model integration technologies, including: Bagging: such as Random Forest, which reduces the variance of the model by combining multiple weak classifiers; Boosting: such as XGBoost and LightGBM, which trains multiple weak classifiers in series, and each classifier corrects the error of the previous classifier, thereby improving the accuracy of the model; Stacking: using the prediction results of different models as input to train a new meta-model for final prediction; model integration can effectively reduce the error of a single model and improve the performance in complex fraud detection tasks.

[0040] Furthermore, the explainability analysis module 400 includes an explainability unit 401, a decision visualization unit 402, a text description generation unit 403, an interaction unit 404 and a feedback unit, the explainability unit 401 is connected to the model training module 300, the decision visualization unit 402 is respectively connected to the explainability unit 401 and the text description generation unit 403, the interaction unit 404 is connected to the text description generation unit 403, and the feedback unit is connected to the interaction unit 404.

[0041] Furthermore, the explainability unit 401 is used to output the decision basis of the fraud detection model; The decision visualization unit 402 is used to display the working mechanism of the fraud detection model in the form of a visualization chart and provide a basis for decision making; The text description generating unit 403 is used to generate a text description to explain the decision of the fraud detection model; The interaction unit 404 is used to provide an interactive interface, record parameter settings selected by different users, and perform comparative analysis to view decision differences between different settings; The feedback unit is used to collect user feedback data on the fraud detection model.

[0042] In this embodiment, the explainability unit 401 outputs the decision basis of the fraud detection model; LIME is used: by approximating the local behavior of the model, an easy-to-understand linear model is generated to show the reason for a decision. For example, LIME can explain why a certain insurance claim is marked as potential fraud. SHAP: Based on the Shapley value of game theory, the contribution of each feature to the model prediction is calculated. SHAP provides global and local explanations, allowing users to understand the decision logic of the model in different situations. Grad-CAM: Mainly used for convolutional neural networks, it can generate heat maps to show which parts of the input data have the greatest impact on the decision, and is suitable for the analysis and interpretation of image data.

[0043] The decision visualization unit 402 uses visual charts to display the working mechanism of the fraud detection model and provide a basis for decision-making. In order to facilitate users to understand the decision-making process of the model, the explainability analysis module 400 generates multiple types of visual charts: Feature importance chart: Displays the influence weight of each feature on the model decision, and is displayed through a bar chart or radar chart to facilitate identification of the factors that have the greatest impact on the final decision. Local explanation chart: Use LIME or SHAP technology to generate explanation charts for specific decisions. For example, it shows the main reasons why a claim is marked as fraudulent and its corresponding feature contribution. Heat map: For images or high-dimensional data, a Grad-CAM heat map is generated to highlight the key areas of concern in the decision-making process. Visual charts enable users to intuitively understand the working mechanism of the model and provide a basis for decision-making. Especially in complex scenarios such as insurance fraud detection, intuitive explanations are crucial. The text description generation unit 403 generates a text description to explain the decision of the fraud detection model; automatically generates a text description to explain the decision of the model in detail: Decision summary: provides a brief overview of the model's judgment results, for example, "The claim is marked as high risk, mainly because the customer's past claim frequency is abnormally high and there are inconsistent social network activity records." Detailed analysis: each key judgment point of the model is analyzed step by step to explain how each feature affects the final decision. For example, "The customer's geographic location is highly overlapped with the location of multiple confirmed fraud cases in the past three months, which increases the risk score." The interactive unit 404 provides an interactive interface, records the parameter settings selected by different users, and conducts comparative analysis to view the decision differences of different settings; the feedback unit collects feedback data for users on the fraud detection model.

[0044] Further, the AI-based insurance fraud prevention and detection system further includes a report generation and feedback module 500, and the report generation and feedback module 500 is connected to the explainability analysis module 400; The report generation and feedback module 500 is used to obtain decision and interaction data, generate fraud detection reports, and receive user feedback data.

[0045] In this embodiment, the report generation and feedback module 500 obtains decision and interaction data, generates a fraud detection report, and receives user feedback data.

[0046] Furthermore, the report generation and feedback module 500 includes a fraud detection report generation unit 501, a dynamic report update unit 502, a report output unit 503, an audit unit 504, a model improvement unit 505 and a user experience optimization unit 506, the fraud detection report generation unit 501 is connected to the explainability analysis module 400, the dynamic report update unit 502 is respectively connected to the fraud detection report generation unit 501 and the report output unit 503, the audit unit 504 is respectively connected to the report output unit 503 and the model improvement unit 505, and the user experience optimization unit 506 is connected to the model improvement unit 505.

[0047] Furthermore, the fraud detection report generating unit 501 is used to summarize the detection results and generate a report using natural language; The dynamic report updating unit 502 is used to receive new data provided by the user and update the generated report; The report output unit 503 is used to output reports in various formats; The review unit 504 is used to manually review high-risk insurance claims; The model improvement unit 505 is used to obtain user feedback data on the fraud detection model and provide data for periodic retraining; The user experience optimization unit 506 is used to optimize the user interaction interface.

[0048] In this embodiment, the fraud detection report generation unit 501 summarizes the detection results and generates a report using natural language to generate a comprehensive report covering the following contents: Summary of detection results: including the fraud risk score of each insurance claim, the marking status such as "high risk" and "low risk", and the recommended further processing measures, for example, "This claim is marked as high risk, and manual review is recommended." Detailed analysis report: In-depth analysis of the components of the fraud risk score, detailing the contribution of each feature in the model. For example, how a customer's historical claim frequency, abnormal geographic activities, social network data, etc. affect the risk score. Explanatory description: Based on the description generated by the explainability analysis module 400, the decision basis and logic of the model are displayed. This part can use feature contribution charts generated by technologies such as LIME and SHAP, or heat maps generated by Grad-CAM. Risk assessment and suggestions: Provide risk assessment for each high-risk case and give response suggestions, such as further investigation, temporary rejection of claims, or other preventive measures. The dynamic report update unit 502 receives new data provided by the user and updates the generated report. When the user provides new evidence or additional customer information, the report will be automatically regenerated and contain the latest analysis results and suggestions to ensure that the report always reflects the most real-time detection results and risk assessments; the report output unit 503 outputs reports in multiple formats, including PDF format, HTML format, and Excel format. The review unit 504 uses manual review of high-risk insurance claims, including: Manual review interface: Users can conduct detailed reviews of high-risk or suspicious insurance claims through the interactive interface, mark misclassifications, or confirm system suggestions. Users can also add comments to record other risk factors or related information found during the review process. Feedback data collection: All user feedback information will be collected by the system and used to update the training data set of the model. The system will use this feedback information to correct the model's misclassification, adjust feature weights, and enhance the detection capability of similar cases in the future. Feedback-driven active learning: Combined with user feedback, the system uses active learning technology to identify weak links in the model and focus on improvement. This process not only improves the accuracy of the model, but also reduces the occurrence of false positives and false negatives. The model improvement unit 505 obtains user feedback data on the fraud detection model and provides data for periodic retraining, and implements model retraining in the following ways: Periodic model retraining: Based on the feedback data, the model is retrained regularly to improve feature extraction and decision logic, thereby enhancing the accuracy of detection. Hyperparameter adjustment: Feedback data helps to identify deficiencies in the current model hyperparameter settings, and automated machine learning will optimize the hyperparameter configuration based on feedback. Introduction of new features: Based on new fraud methods or patterns discovered through user feedback, the system can introduce new features for modeling to address future fraud risks; the user experience optimization unit 506 optimizes the user interaction interface.

[0049] What is disclosed above is only one or more preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of the present invention still fall within the scope of the invention.

Claims

1. An AI-based insurance fraud prevention and detection system, characterized in that: It includes a data collection module, a data preprocessing module, a model training module and an interpretability analysis module, wherein the data preprocessing module is connected to the data collection module, and the model training module is connected to the data preprocessing module and the interpretability analysis module respectively; The data collection module is used to obtain data from various data sources and integrate them after processing; The data preprocessing module is used to preprocess the integrated data to obtain normalized data; The model training module is used to build a fraud detection model and make decisions based on the preprocessed data; The explainability analysis module is used to display decision results and interact with users.

2. The AI-based insurance fraud prevention and detection system according to claim 1, characterized in that: The data collection module includes a multi-source data acquisition unit, a preprocessing unit, a feature extraction unit, a data storage unit, a data protection unit and a real-time data processing unit. The preprocessing unit is respectively connected to the multi-source data acquisition unit and the feature extraction unit, the data storage unit is connected to the feature extraction unit, and the data protection unit is respectively connected to the data storage unit and the real-time data processing unit.

3. The AI-based insurance fraud prevention and detection system according to claim 2, characterized in that: The multi-source data acquisition unit is used to collect data from multiple heterogeneous data sources; The pre-processing unit is used to clean, convert and enhance the collected data; The feature extraction unit is used to extract features from complex relational data and perform data integration; The data storage unit is used to store the integrated data; The data protection unit is used to encrypt the transmission data; The real-time data processing unit is used to integrate the data processing framework and process data in real time.

4. The AI-based insurance fraud prevention and detection system according to claim 1, characterized in that: The data preprocessing module includes a data cleaning unit, a feature selection unit and a data segmentation unit. The data cleaning unit is connected to the data collection module, and the feature selection unit is connected to the data cleaning unit and the data segmentation unit respectively.

5. The AI-based insurance fraud prevention and detection system according to claim 4, characterized in that: The data cleaning unit is used to remove noise and outliers from the data; The feature selection unit is used to select features to obtain a data set, and to expand and transform the data set to obtain a standardized data set; The data segmentation unit is used to segment the data set into a training set, a validation set and a test set.

6. The AI-based insurance fraud prevention and detection system according to claim 5, characterized in that: The model training module includes a deep learning model training unit, a graph neural network model, a model updating unit, an optimization unit and a model integration unit. The deep learning model training unit is connected to the data preprocessing module, the graph neural network model is respectively connected to the deep learning model training unit and the model updating unit, and the optimization unit is respectively connected to the model updating unit and the model integration unit.

7. The AI-based insurance fraud prevention and detection system according to claim 6, characterized in that: The deep learning model training unit is used to provide a deep learning model and use the acquired training set, validation set and test set to perform initial training on the fraud detection model; The graph neural network model is used to provide a graph neural network model, modeling through nodes and edges, capturing high-order relationship features in the data, performing secondary training on the fraud detection model, and making decisions based on the data; The model updating unit is used to collect update data, annotate the update data, and periodically retrain the fraud detection model using the update data; The optimization unit is used to optimize the parameter configuration of the fraud detection model; The model integration unit is used to perform multi-model integration on the fraud detection model.

8. The AI-based insurance fraud prevention and detection system according to claim 1, characterized in that: The explainability analysis module includes an explainability unit, a decision visualization unit, a text description generation unit, an interaction unit and a feedback unit. The explainability unit is connected to the model training module, the decision visualization unit is respectively connected to the explainability unit and the text description generation unit, the interaction unit is connected to the text description generation unit, and the feedback unit is connected to the interaction unit.

9. The AI-based insurance fraud prevention and detection system according to claim 8, characterized in that: The explainability unit is used to output the decision basis of the fraud detection model; The decision visualization unit is used to display the working mechanism of the fraud detection model in the form of a visualization chart and provide a basis for decision-making; The text description generating unit is used to generate a text description to explain the decision of the fraud detection model; The interactive unit is used to provide an interactive interface, record parameter settings selected by different users, and conduct comparative analysis to view decision differences between different settings; The feedback unit is used to collect user feedback data on the fraud detection model.

Images