An auto insurance data optimized retrieval method for a digital management platform

Abstract

The application relates to the technical field of data retrieval and fraud detection, and discloses a vehicle insurance data optimization retrieval method for a digital management platform, which comprises the following steps: acquiring vehicle insurance data, performing multi-view feature extraction, and constructing a unified feature space representation; generating high-quality fraud samples by using a conditional generative adversarial network to enhance model training; constructing a deep reconstruction network to amplify abnormal feature signals through reconstruction error; constructing a difference-sensitive hash index structure based on the features of the reconstruction error; comprehensively evaluating the retrieved suspected fraud samples by using a multi-view consistency verification method to output final fraud detection results; and generating high-quality fraud samples by using the conditional generative adversarial network, so that the problem of data imbalance caused by the scarcity of fraud samples is effectively solved, and the model generalization capability is improved.

Description

Technical Field

[0001] This invention relates to the fields of data retrieval and fraud detection technology, and more specifically, to an optimized retrieval method for vehicle insurance data for digital management platforms. Background Technology

[0002] As the digital transformation of the auto insurance industry deepens, insurance companies have accumulated massive amounts of multimodal data, including policy information, claims records, accident images, and text descriptions. Efficiently and accurately retrieving and identifying fraud cases from this data is crucial for reducing operating costs for insurance companies and maintaining market fairness.

[0003] Currently, auto insurance fraud detection mainly relies on manual review and simple rule matching, which suffers from high workload and low efficiency. With the development of machine learning technology, some insurance companies have tried to use supervised learning models for fraud detection, but they face three core technical challenges: auto insurance fraud cases account for a very small percentage of the overall data (usually less than 1% of the total sample), and the severe data imbalance makes model training difficult; fraud features are often deliberately hidden or disguised, and conventional feature extraction methods are unable to identify hidden abnormal patterns; fraud methods are constantly being updated, while high-quality fraud samples are scarce, resulting in weak generalization ability of the detection model.

[0004] Existing anomaly detection methods for multimodal data mainly include: statistical analysis-based anomaly detection, which is only applicable to simple data distributions and has limited ability to extract complex features; unsupervised anomaly detection based on isolated forests, which is highly dependent on feature engineering and lacks multimodal fusion capabilities; representation learning methods based on deep learning, which lack effective learning mechanisms for small sample fraud categories; and similarity retrieval methods based on locality-sensitive hashing, which are not optimized for fraud features and have insufficient retrieval accuracy.

[0005] Existing technologies lack a comprehensive solution that simultaneously addresses the three technical challenges of data imbalance, feature concealment, and efficient retrieval. Summary of the Invention

[0006] This invention provides an optimized retrieval method for vehicle insurance data for digital management platforms, solving the technical problems of data imbalance, concealed fraud features, and low retrieval efficiency in related technologies.

[0007] This invention provides a method for optimizing and retrieving auto insurance data for digital management platforms, comprising: Acquire vehicle insurance data, perform multi-view feature extraction on the vehicle insurance data, and construct a unified feature space representation; By utilizing feature space representation, a conditional generative adversarial network is used to generate high-quality fraud samples, enhance model training, and generate an enhanced dataset. Based on feature space representation and augmented datasets, a deep reconstruction network is constructed to amplify abnormal feature signals through reconstruction errors, thereby enhancing the detection effect of fraudulent samples. Based on the characteristics of reconstruction error, a difference-sensitive hash index structure is constructed to achieve efficient fraud sample retrieval; For suspected fraudulent samples retrieved, multi-view analysis is performed. Figure 1 The consistency verification methods are comprehensively evaluated, and the final fraud detection result is output.

[0008] Furthermore, the multi-view feature extraction step includes: Feature engineering methods are used to extract structured features from vehicle insurance data and generate structured feature vectors. The text feature extraction model is used to process the text data of the accident and generate text feature vectors. Image feature extraction models are used to process image data from accident scenes and generate image feature vectors. The feature vectors of each modality are mapped to a unified feature space to generate a comprehensive feature representation.

[0009] Furthermore, the text feature extraction employs a pre-trained language model, and the calculation method is as follows: ; in For text feature vectors, This indicates the input text. Indicates a text encoder; The image feature extraction employs a convolutional neural network model, and the calculation method is as follows: ; in For image feature vectors, Indicates the input image. This indicates an image encoder.

[0010] Furthermore, the step of generating high-quality fraud samples using the conditional generative adversarial network includes: Construct a conditional generative adversarial network structure that includes a generator G and a discriminator D; The parameters of generator G and discriminator D are optimized through an adversarial training process that minimizes generator loss and maximizes discriminator loss. The generated fraudulent samples are evaluated and screened for quality, and high-quality samples are retained for subsequent model training.

[0011] Furthermore, the objective function of the adversarial training process is: ; in Indicates the generator The parameter minimization operation, Indicates the discriminator Parameter maximization operation Represents the value function. Indicates the expected value. This represents the log probability of the discriminator on the true sample. Represents the true sample distribution. This represents the complement of the logarithmic probability of the discriminator for the generated samples. Represents a random noise vector. Represents a condition vector. Represents a generator function; The sample quality assessment employs a method combining the output probability of discriminator D and a feature space distance metric, calculated as follows: ; in Indicates the generation of samples quality The weighting coefficients represent the output probability of the discriminator. This indicates that the discriminator evaluates the generated samples. The output probability, Indicates the generation of samples Compared with real fraud sample sets similarity, This represents the condition vector.

[0012] Furthermore, the step of constructing the deep reconstruction network includes: Construct a deep reconstruction network structure that includes an encoder and a decoder; The deep reconstruction network is trained by minimizing the reconstruction loss function, enabling it to accurately reconstruct the features of normal samples. For the sample to be detected, its reconstructed sample is calculated, and then a reconstruction error map is generated; Anomaly scores are calculated for samples based on reconstruction error maps, serving as an indicator of the likelihood of fraud.

[0013] Furthermore, the reconstruction loss function is defined as: ; in Indicates the losses incurred during reconstruction. This represents the reconstruction error of the sample space. Indicates sample features, Represents the reconstructed sample features. Denotes the square of the L2 norm. For feature extraction function, Indicates from original features The high-level feature representation extracted from it, Indicates from reconstruction features The high-level feature representation extracted from it, This represents the reconstruction error of the feature space. Weighting coefficients to balance the errors of the two parts; The abnormal score is calculated as follows: ; in Indicates sample Abnormal scores, Representing feature dimension, The first part represents the reconstruction error map. One element, Indicates the first The importance weights of each feature This indicates a summation operation across all feature dimensions.

[0014] Furthermore, the difference-sensitive hash index construction step includes: The SMOTE variant algorithm is used to balance the dataset and increase the proportion of minority class samples. Based on the principle of locality-sensitive hashing, a set of hash functions is constructed to map the high-dimensional feature space to the low-dimensional hash code space. Enhance the difference sensitivity of locality-sensitive hash functions to improve their sensitivity to fraudulent features; A multi-level index structure is built based on hash encoding, which supports fast similarity retrieval.

[0015] Furthermore, the hash function is calculated as follows: ; in Representing the eigenvector One-dimensional hash code, It is a symbolic function. It is a random projection vector that follows a standard normal distribution. ; Points representing the projection vector and eigenvector; It is a random bias term that follows a uniform distribution. ; The enhanced difference sensitivity is achieved by weighting the projection vector according to the feature importance: ; in Represents element-level multiplication. Represents the original random projection vector. It is a feature importance vector. It is an adjustment factor. This represents the adjusted projection vector.

[0016] This invention provides a vehicle insurance data optimization retrieval system for digital management platforms, used to execute the aforementioned vehicle insurance data optimization retrieval method for digital management platforms, comprising: The multi-view feature extraction module is used to extract multimodal features from vehicle insurance data and map the extracted features to a unified feature space. The sample augmentation module is used to generate fraudulent samples based on the principle of conditional generative adversarial networks, balance the data distribution, and improve the model's generalization ability. The anomaly feature enhancement module is used to calculate the reconstruction error through a deep reconstruction network, amplify fraud feature signals, and generate anomaly scores. The high-efficiency indexing module is used to build differentially sensitive hash indexes based on anomaly features, enabling fast retrieval of fraudulent samples; The multi-view verification module is used to collect fraud evidence from multiple perspectives, assess consistency, calculate suspiciousness scores, and output the final detection results.

[0017] The beneficial effects of this invention are as follows: It generates high-quality fraud samples through conditional generative adversarial networks, effectively solving the data imbalance problem caused by the scarcity of fraud samples and improving the model's generalization ability; it enhances the abnormal features and multi-view capabilities of deep reconstruction networks. Figure 1 Consistency verification can effectively detect deliberately hidden or disguised fraudulent features, improving detection accuracy; efficient retrieval based on difference-sensitive hash index improves retrieval speed while maintaining high retrieval accuracy; the system can be directly integrated into existing business systems without additional hardware investment, reducing implementation costs. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for optimizing and retrieving vehicle insurance data for a digital management platform, as described in this invention. Figure 2 This is a flowchart of step 1 of the present invention; Figure 3 This is a flowchart of step 2 of the present invention; Figure 4 This is a flowchart of step 3 of the present invention; Figure 5 This is a flowchart of step 4 of the present invention; Figure 6 This is a flowchart of step 5 of the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses an optimized retrieval method for vehicle insurance data for digital management platforms, such as... Figures 1 to 6 As shown, it includes: Step 1: Obtain vehicle insurance data, extract multi-view features from the vehicle insurance data, and construct a unified feature space representation; Multi-view feature extraction is performed on auto insurance data to construct a unified feature space representation, specifically including: Step 1.1, Structured data feature extraction; Feature engineering methods are used to extract structured data features from auto insurance, including numerical features such as claim amount, frequency, and time interval, as well as categorical features such as insurance type and vehicle information, to generate structured feature vectors.

[0021] For numerical features, standardization is applied: ; in Represents the standardized eigenvalues. Represents the original feature values. Represents the characteristic mean. It represents the characteristic standard deviation.

[0022] For categorical features, One-Hot encoding is used to convert them into numerical representations.

[0023] Specifically, for those with The categorical features with 10 possible values ​​will be transformed into... A binary feature vector of dimension 1, where only one value is 1 and the rest are 0, corresponding to the category of the original feature.

[0024] Step 1.2, Text Description Feature Extraction; The text feature extraction model is used to process text data such as accident descriptions and police reports to generate text feature vectors. The text feature extraction employs a pre-trained language model to map the input text into fixed-dimensional feature vectors. ; in For text feature vectors, This indicates the input text. Indicates a text encoder; The text encoder is the BERT encoder, whose architecture includes a multi-layer bidirectional Transformer encoder. After the input text is segmented, embedded, and positionally encoded, it is processed through a multi-head self-attention mechanism and a feedforward neural network layer to finally extract the semantic representation of the text.

[0025] Step 1.3, Image Feature Extraction; Image feature extraction models are used to process image data such as accident scene photos to generate image feature vectors. The image feature extraction employs a convolutional neural network model to map the input image into a fixed-dimensional feature vector. ; in For image feature vectors, Indicates the input image. Indicates an image encoder; The image encoder is a ResNet-50 model, which consists of 50 deep residual networks, including convolutional layers, batch normalization layers, activation function layers, and residual connections. It can effectively extract hierarchical features of images. The image is first preprocessed (resized and normalized) and then input into the ResNet-50 model. The 2048-dimensional feature vector output from the last pooling layer is extracted as the image feature representation.

[0026] Step 1.4, Multimodal feature fusion; The feature vectors of each modality are mapped to a unified feature space to generate a comprehensive feature representation: ; in To fuse feature vectors, , , These represent the weight coefficients for structured features, text features, and image features, respectively. The initial values ​​of the three weight coefficients can be set to be equal, and then optimized through model training.

[0027] In practice, a multilayer perceptron (MLP) network is used to map and fuse the features of each modality. The specific structure is as follows: first, an independent fully connected layer is applied to each modality feature for dimensionality reduction mapping, then the mapped features are concatenated, and finally, a common fully connected layer is used to generate the final fused features.

[0028] Step 2: Utilize feature space representation and employ conditional generative adversarial networks to generate high-quality fraud samples, enhance model training, and generate an enhanced dataset. To address the problem of scarce fraud samples, Conditional Generative Adversarial Networks (CGANs) are used to generate high-quality fraud samples, enhancing model training. Specifically, this includes: Step 2.1: Construct a conditional generative adversarial network; This step utilizes the multi-view features extracted in step 1. As the basic data for network training, it is used as a discriminator. The input is used to distinguish between real samples and generated samples.

[0029] Build containing generators and discriminator Conditional generative adversarial network structure; generator Learning from random noise and conditions (e.g., fraud type, feature distribution, etc.) Generate realistic fraud samples and a discriminator. It then learns to distinguish between real and generated samples.

[0030] generator The specific structure is a multi-layer fully connected neural network: the input layer receives a random noise vector. (Dimension 100) and condition vector (Dimension is the number of conditional categories), containing three hidden layers (with 256, 512, and 1024 nodes respectively). Each hidden layer is followed by a LeakyReLU activation function and a batch normalization layer. The output layer uses the Tanh activation function to generate synthetic samples with the same dimension as the features.

[0031] generator The input is a random noise vector and a conditional vector, and the output is the synthesized fraud sample features: ; in This represents the features of the generated synthetic fraud samples. Represents a generator function. Represents a random noise vector. This represents the condition vector.

[0032] Discriminator The specific structure is also a multi-layer fully connected neural network: the input layer receives sample features. and condition vector It contains three hidden layers (with 1024, 512, and 256 nodes respectively), each followed by a LeakyReLU activation function and a Dropout layer (with a dropout rate of 0.3). The output layer uses a Sigmoid activation function to generate a probability value.

[0033] Discriminator The inputs are sample features (real or generated samples) and a condition vector, and the output is the probability value of the sample's authenticity. ; in This represents the discriminator function. Indicates sample features, Represents a condition vector. Representing the features of a given sample and conditions When, the probability that the sample is a real sample.

[0034] Step 2.2, the adversarial training process; The parameters of the generator G and discriminator D are optimized through an adversarial training process that minimizes the generator loss and maximizes the discriminator loss. The objective function is: ; in Indicates the generator The parameter minimization operation, Indicates the discriminator Parameter maximization operation Represents the value function. Indicates the expected value. This represents the log-probability of the discriminator with respect to the true sample. Represents the true sample distribution. This represents the complement of the logarithmic probability of the discriminator for the generated samples. Represents a random noise vector. Represents a condition vector. This represents a generator function.

[0035] During training, the discriminator and generator Alternately update parameters: The parameter update objective of discriminator D is to maximize That is, to improve the ability to distinguish between real samples and generated samples; The parameter update objective of generator G is to minimize That is, to improve the ability to generate realistic samples; The specific process of adversarial training is as follows: First, fix the parameters of the generator G and train the discriminator D in multiple batches; then fix the parameters of the discriminator D and train the generator G in one batch. This process is repeated until the generator can generate high-quality fraudulent samples, and the discriminator has difficulty distinguishing between real samples and generated samples (the discriminator accuracy is close to 50%).

[0036] Step 2.3, Sample Screening and Enhancement; The generated fraudulent samples undergo quality assessment and screening, retaining high-quality samples for subsequent model training. The quality assessment employs a method combining the output probability of the discriminator D and a feature space distance metric. ; in Indicates the generation of samples quality The weighting coefficients represent the output probability of the discriminator. This indicates that the discriminator evaluates the generated samples. The output probability, Represents a condition vector. Indicates the generation of samples Compared with real fraud sample sets The similarity is calculated using cosine similarity: ; in This represents the number of samples in the real fraud sample set. This represents a sample from the set of real fraud samples. Indicates sample and dot product, Indicates sample L2 norm, Indicates sample The L2 norm.

[0037] The selected high-quality samples, together with the original fraudulent samples, constitute the augmented dataset. Used for subsequent model training; the selection criteria are: ; in The quality threshold is set to 0.75 in this example.

[0038] Step 3: Based on the feature space representation and augmented dataset, construct a deep reconstruction network to amplify abnormal feature signals through reconstruction errors, thereby enhancing the detection effect of fraudulent samples; A deep reconstruction network is constructed to amplify abnormal feature signals through reconstruction errors, thereby enhancing the detection effect of fraudulent samples. Specifically, this includes: Step 3.1, construct a deep reconstruction network; This step is based on the multi-view fusion feature output in step 1. and the augmented dataset generated in step 2 A deep reconstruction network is constructed. This network can accurately reconstruct normal samples by learning the distribution characteristics of normal samples, while producing a large reconstruction error for fraudulent samples.

[0039] A deep reconstruction network structure containing an encoder and a decoder is constructed. Both the encoder and decoder are implemented using multi-layer fully connected neural networks.

[0040] Encoder structure: The input layer receives feature vectors (with the same dimension as the original features), and then the dimensionality is progressively reduced through three fully connected layers (512, 256, and 128 nodes respectively). Each layer uses the ReLU activation function and batch normalization. The encoder processes the input samples... Compression into latent representation (Dimensions are 128): ; in Represents the encoder function. Indicates sample features, This represents the latent representation vector.

[0041] Decoder structure: The input layer receives the latent representation. Then, the dimension is gradually restored through three fully connected layers (with 256, 512 nodes and the original feature dimension, respectively). The intermediate layers use ReLU activation function and batch normalization, and the last layer uses activation function that matches the input feature value domain (such as Sigmoid or Tanh).

[0042] The decoder will Reconstructed : ; in Represents the decoder function. Represents the latent representation vector. This represents the reconstructed sample features.

[0043] Overall, the structure of deep reconstruction networks resembles an hourglass, with the underlying representation in the middle... It is an information bottleneck that forces networks to learn the most important characteristics of data.

[0044] Step 3.2, Optimize reconstruction loss; A deep reconstruction network is trained by minimizing the reconstruction loss function, enabling it to accurately reconstruct features from normal samples. The reconstruction loss function is defined as: ; in Indicates the losses incurred during reconstruction. This represents the reconstruction error of the sample space. Indicates sample features, Represents the reconstructed sample features. Denotes the square of the L2 norm. For feature extraction function, Indicates from original features The high-level feature representation extracted from it, Indicates from reconstruction features The high-level feature representation extracted from it, This represents the reconstruction error of the feature space. The weighting coefficients are used to balance the errors of the two parts.

[0045] The network was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 training epochs. Normal samples were primarily used during training to allow the reconstruction network to learn their distribution characteristics. For fraudulent samples, the reconstruction process would introduce significant errors due to their different feature distribution compared to normal samples.

[0046] Step 3.3, Reconstruction of error map generation; For the sample to be tested Calculate its reconstructed sample Then, a reconstruction error map is generated: ; in This represents element-level absolute value operations. Indicates sample features, Represents the reconstructed sample features. The reconstruction error map is a vector with the same dimensions as the original features, where each element represents the reconstruction error of the corresponding feature.

[0047] For better visualization, the reconstruction error map can be displayed as a heatmap, with the color intensity corresponding to the error magnitude, helping analysts to intuitively identify abnormal feature areas.

[0048] Step 3.4, Calculation of outlier scores; Anomaly scores are calculated for samples based on reconstruction error maps, serving as an indicator of the likelihood of fraud. ; in Indicates sample Abnormal scores, Representing feature dimension, The first part represents the reconstruction error map. One element, Indicates the first The importance weights of each feature This indicates a summation operation across all feature dimensions.

[0049] Feature importance weights The model is trained using a gradient boosting tree model, which uses known fraudulent / normal sample labels as targets, features as inputs, and then extracts the importance score of each feature as a weight.

[0050] Step 4: Based on the characteristics of reconstruction error, construct a difference-sensitive hash index structure to achieve efficient fraud sample retrieval; Based on the dataset after sample balancing, a difference-sensitive hash index structure is constructed to achieve efficient fraud sample retrieval, specifically including: Step 4.1, sample balancing; This step utilizes the augmented dataset generated in step 2. As a foundation, and in conjunction with the feature importance weights obtained in step 3. Perform sample balancing to further improve data quality.

[0051] A variant of the Synthetic Minority Over-sampling Technique (SMOTE) algorithm is used to balance the dataset, increasing the proportion of minority class samples (fraudulent samples). The basic idea is to generate synthetic samples in the feature space between the minority class samples. For each minority class sample Randomly select a sample from its k nearest neighbors. Then and Linear interpolation between them generates new samples : ; in Indicates minority class samples, express A k-nearest neighbor sample, Indicates the interpolation coefficients. This represents the difference vector between two samples. This indicates the newly generated sample.

[0052] To adapt to the specific needs of fraud detection, the standard SMOTE algorithm was improved by introducing anomaly weight adjustment and feature importance weighting methods to generate more realistic synthetic fraud samples. Specific improvements include: Nearest neighbor selection optimization: Instead of randomly selecting a sample from the k nearest neighbors, a weighted selection is made based on the sample quality score (defined by the outlier score in the previous steps), and samples with higher quality scores are more likely to be selected; Feature interpolation weighting: interpolation coefficients for different features Unlike other methods, a smaller range of variation is used for important features (determined through feature importance analysis) to maintain the integrity of fraud features; Noise addition: Add small-amplitude Gaussian noise to the generated synthetic samples to increase sample diversity; Through the above improvements, the SMOTE variant algorithm can generate more realistic and diverse fraud samples, thereby improving the generalization ability of subsequent models.

[0053] Step 4.2, Construction of Locality Sensitive Hash Function; Based on the principle of Locality Sensitive Hash (LSH), a set of hash functions is constructed to map the high-dimensional feature space to the low-dimensional hash code space, while maintaining the similarity of similar samples in the feature space in the hash code space.

[0054] For feature vectors Calculate its hash code : ; in It is a random projection vector that follows a standard normal distribution. ; It is a random bias term that follows a uniform distribution. , It is a symbolic function. This represents the dot product of the projection vector and the eigenvector. Representing the eigenvector A hash code.

[0055] By using multiple independent hash functions: ; in , , They represent the 1st, 2nd, and 3rd respectively. Each independent hash function This represents the total number of hash functions.

[0056] Each hash function contributes one bit: ; in Representing the eigenvector The complete hash code of is an m-dimensional vector. This represents a vector concatenation operation. , , They represent the 1st, 2nd, and 3rd respectively. Each hash function applies to the feature vector The hash result, This represents the total number of hash functions.

[0057] In this embodiment, the hash code length Setting it to 128 provides a good balance between feature representation capability and retrieval efficiency.

[0058] Step 4.3, enhanced sensitivity to differences; To improve sensitivity to fraud features, the locality-sensitive hash function was enhanced with difference sensitivity. This step directly utilizes the feature importance weights obtained in step 3. As a basis for enhanced difference sensitivity, hash functions become more sensitive to important features.

[0059] For the projection vector Introduce feature importance weighted adjustment: ; in Represents element-level multiplication. Represents the original random projection vector. It is a feature importance vector. This is an adjustment factor, set to 0.8 in this embodiment. This represents the adjusted projection vector.

[0060] Feature importance can be obtained by analyzing the feature discrimination between fraudulent samples and normal samples, specifically calculated using the Information Gain Ratio: ; in, It is a feature Information gain, representing the use of features Reduced information entropy during classification; Indicates the first One feature is used to distinguish between fraudulent and legitimate samples; It is a feature The intrinsic value of is used to standardize information gain, and the calculation formula is: ; in It is a feature Take the first A subset of samples with possible values, Representing a subset The number of samples in It is the total number of samples. Representation of features The number of possible values, Representation of features Take the first The probability of each possible value.

[0061] Feature importance vector The IGR values ​​of each feature are obtained by normalizing them to the [0,1] interval. The normalization formula is as follows: ; in Indicates the first Normalized importance values ​​of each feature This represents the minimum information gain ratio among all features. This represents the maximum information gain ratio among all features, ensuring that the importance values ​​of all features are compared on a uniform scale.

[0062] In this way, the difference-sensitive hash function is more sensitive to the key features of fraudulent samples and can better distinguish between fraudulent samples and normal samples.

[0063] Step 4.4, Hash index structure construction; A multi-level index structure based on hash encoding is constructed to support fast similarity retrieval. The index structure uses key-value storage, where the key is a hash code and the value is the corresponding sample set.

[0064] To handle hash collisions and improve retrieval accuracy, a multi-table hashing strategy is adopted to construct... A separate hash table: ; ; in Indicates the first A hash table, It is the first The set of hash functions used by each hash table It is the first One sample, It is the total number of samples. This represents key-value pairs in a hash table. It is a sample In the Hash codes in hash tables Indicates the number of hash tables.

[0065] In this embodiment, the number of hash tables Set to 10, each hash table uses a different set of random projection vectors, providing a multi-angle view of the feature map.

[0066] The query process for the index structure is as follows: For the query sample Calculate its hash code in each hash table. Then, samples with the same or similar hash codes are searched in each hash table, the results are merged and sorted by similarity, and the k most similar samples are returned. To handle hash coding errors, a small number of bit differences in hash codes can be allowed (e.g., Hamming distance not exceeding 3).

[0067] Step 5: For the suspected fraud samples retrieved, use multi-view... Figure 1 The consistency verification methods are comprehensively evaluated, and the final fraud detection result is output. For suspected fraudulent samples retrieved, multi-view analysis is performed. Figure 1 The consistency verification methods are comprehensively evaluated, and the final fraud detection result is output, including: Step 5.1, Multi-view evidence collection; This step comprehensively utilizes the output results of the previous steps to collect fraud evidence from multiple different views, enabling multi-angle judgment and improving detection accuracy. It mainly includes the following three views: Reconstructed view: Reconstruction error map based on the output of the deep reconstruction network in step 3. and abnormal scores ; Similarity view: Similar fraud cases retrieved based on the difference-sensitive hash index in step 4; Statistical view: Statistical distribution anomaly based on each feature of the sample; For reconstructed views, fraud evidence score Directly use abnormal scores Normalization is performed.

[0068] For similarity views, fraud evidence scores Fraud rate calculation based on retrieved similar samples: ; in Indicates sample Fraud evidence score of similarity view It is the first one retrieved 10 similar samples It is a sample The label (1 indicates fraud, 0 indicates normal). It is a sample and similarity, This indicates the number of similar samples considered. Indicates the previous Summing is performed on 10 similar samples.

[0069] For the statistical view, fraud evidence score Calculation based on the degree of anomaly of each feature value in the feature distribution of the sample: ; in Indicates sample Fraud evidence score in the statistical view. It is the feature dimension. It is a feature The Z-score represents the number of standard deviations from the characteristic mean. This indicates summing over all feature dimensions.

[0070] For each view Calculate its fraud evidence score , indicating sample The likelihood of fraud as seen from this view. All evidence scores have been normalized to ensure a value range of [0,1].

[0071] Step 5.2, Evidence Consistency Assessment; Assess the consistency of evidence across different views and calculate the consistency coefficient. : ; in Indicates sample Consistency coefficient, Represents the standard deviation function. Represents the average function, Represents a view. It is the total number of views. Indicates sample In view Fraud evidence score. Consistency coefficient. The higher the value, the more consistent the evidence from different views; When the consistency coefficient When the value is negative, it is truncated to 0.

[0072] The consistency coefficient is used as a weighting factor in the subsequent calculation of the suspicion score, which improves the suspicion score of samples with high consistency of evidence from different views.

[0073] Step 5.3, Suspicion score calculation; Taking into account multiple view evidence and consistency, the suspiciousness score of the sample is calculated: ; in Indicates sample Suspicion rating, Indicates sample With category The distance is calculated using Mahalanobis distance: ; in It is a category The mean vector, It is a category The covariance matrix, Denotes the inverse matrix of the covariance matrix. Representing vectors transpose, This represents the square root operation.

[0074] Indicates the first Fraud categories, Indicates the first One normal category, This represents the total number of fraud categories. This represents the total number of normal categories. This represents the summation over all fraud categories. This represents the summation over all normal categories. and These represent the weight coefficients for the fraud category and the normal category, respectively, and the consistency coefficient. The relevant calculation formula is as follows: ; ; in and These represent different categories of fraud. and normal category The basic weights, Indicates sample Consistency coefficient, It is the consistency impact coefficient, which controls the degree of influence of consistency on weight adjustment; The adjustment factor representing the weight of the fraud category. This represents the adjustment factor for the weight of the normal category.

[0075] Suspicion rating The higher the value, the better the sample size. The greater the likelihood of fraud.

[0076] Step 5.4, Threshold Decision and Result Output; Based on suspicion level score and preset threshold Make a judgment and output the final fraud detection result: ; in Indicates sample The fraud detection result is 1, indicating fraud, and 0, indicating normal. Indicates sample Suspicion rating, Indicates the decision threshold. This indicates a conditional judgment. Indicates other situations.

[0077] threshold The threshold can be dynamically adjusted according to business needs to achieve a balance between detection rate and false alarm rate. In this embodiment, the optimal threshold is selected through ROC curve analysis. .

[0078] Step 5 is the final step in the entire method, integrating the processing results of all previous steps and outputting the final fraud detection conclusion. Its main outputs include: Suspicion rating : Quantify the likelihood of fraud in a sample.

[0079] Final fraud detection results The result of the binary judgment is 1, which indicates fraud, and 0 indicates normal.

[0080] Multi-view evidence visualization: helps analysts understand the basis for fraud determination and improves the interpretability of results.

[0081] These outputs directly serve business decisions and can be integrated with the auto insurance claims management system to support application scenarios such as fraud risk warning, case prioritization, and claims process optimization.

[0082] A vehicle insurance data optimization retrieval system for a digital management platform, used to execute the aforementioned vehicle insurance data optimization retrieval method for a digital management platform, includes: The multi-view feature extraction module is used to extract multimodal features from vehicle insurance data and map the extracted features to a unified feature space. The sample augmentation module is used to generate fraudulent samples based on the principle of conditional generative adversarial networks, balance the data distribution, and improve the model's generalization ability. The anomaly feature enhancement module is used to calculate the reconstruction error through a deep reconstruction network, amplify fraud feature signals, and generate anomaly scores. The high-efficiency indexing module is used to build differentially sensitive hash indexes based on anomaly features, enabling fast retrieval of fraudulent samples; The multi-view verification module is used to collect fraud evidence from multiple perspectives, assess consistency, calculate suspiciousness scores, and output the final detection results.

[0083] Here, the present invention provides an implementation example: This implementation method has been successfully applied to the auto insurance fraud detection system of a large insurance company. The implementation process and effects of the key steps are described in detail below in combination with the actual application scenario.

[0084] A large insurance company processes approximately 50,000 auto insurance claims monthly, with fraud cases accounting for about 0.8%. Traditional detection methods rely on manual review and simple rule matching, which suffers from high false negative rates, low detection efficiency, and difficulty in combating new fraud tactics. The company possesses approximately 3 million claim records accumulated over the past five years, including structured data (insurance information, claim amounts, etc.), text data (accident descriptions, report records, etc.), and image data (accident scene photos, vehicle damage photos, etc.).

[0085] Multimodal feature extraction and fusion were performed on the company's existing auto insurance data to obtain a unified feature representation. Examples of key structured feature extraction are shown in Table 1. Table 1: Examples of Main Structured Feature Extraction

[0086] For text data, a pre-trained BERT model is used to extract features. Taking an accident description as an example, the text "Driving on the highway at night, the car in front suddenly braked, causing a rear-end collision, and my vehicle's front bumper and headlights were damaged" is input into the BERT model, resulting in a 768-dimensional text feature vector.

[0087] For image data, a pre-trained ResNet-50 model is used to extract features. Preprocessed accident scene photos are input into the model, and the final pooling layer outputs a 2048-dimensional feature vector.

[0088] Multimodal feature fusion is achieved through a multilayer perceptron network, which maps structured features (128-dimensional), text features (768-dimensional), and image features (2048-dimensional) to a unified 256-dimensional feature space to obtain the final fused feature representation.

[0089] Based on 2,400 fraud samples identified from the company's historical data, a conditional generative adversarial network (CGAN) was used to generate high-quality synthetic fraud samples to address the imbalanced sample problem. The CGAN training parameters and generation results are shown in Table 2. Table 2: CGAN Training Parameters and Generation Results

[0090] A deep reconstruction network is constructed to amplify the abnormal feature signals of fraudulent samples by exploiting reconstruction errors. The reconstruction network structure and performance are compared in Table 3. Table 3: Comparison of Reconstructed Network Structure and Results

[0091] Based on the dataset after sample balancing, a difference-sensitive hash index structure is constructed to achieve efficient fraud sample retrieval. The hash index parameters and the importance of key features are shown in Table 4. Table 4: Importance of Hash Index Parameters and Key Features

[0092] For suspected fraudulent samples retrieved, multi-view analysis is performed. Figure 1 The consistency verification methods are comprehensively evaluated, and the final fraud detection result is output. Table 5 shows the multi-view evaluation results and overall system performance of a typical case: Table 5: Multi-view evaluation results and overall system performance of typical cases

[0093] Table 6 shows a comparison of system performance and economic benefits before and after implementation. Table 6: Performance Comparison and Economic Benefits Before and After System Implementation

[0094] The long-term operating indicators of the system are shown in Table 7: Table 7: Long-term operating indicators of the system

[0095] Through this implementation method, the insurance company successfully built a highly efficient auto insurance data optimization and retrieval system. This system not only significantly improved the accuracy and efficiency of fraud detection but also optimized the overall claims process and enhanced customer satisfaction. The implementation of the system also reduced the workload of manual review, allowing reviewers to focus on handling high-risk cases marked by the system, thus improving overall work efficiency.

[0096] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for optimizing and retrieving vehicle insurance data for a digital management platform, characterized in that, include: Acquire vehicle insurance data, perform multi-view feature extraction on the vehicle insurance data, and construct a unified feature space representation; By utilizing feature space representation, a conditional generative adversarial network is used to generate high-quality fraud samples, enhance model training, and generate an enhanced dataset. Based on feature space representation and augmented datasets, a deep reconstruction network is constructed to amplify abnormal feature signals through reconstruction errors, thereby enhancing the detection effect of fraudulent samples. Based on the characteristics of reconstruction error, a difference-sensitive hash index structure is constructed to achieve efficient fraud sample retrieval; For the suspected fraud samples retrieved, a comprehensive evaluation is performed using a multi-view consistency verification method, and the final fraud detection result is output.

2. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 1, characterized in that, The multi-view feature extraction steps include: Feature engineering methods are used to extract structured features from vehicle insurance data and generate structured feature vectors. The text feature extraction model is used to process the text data of the accident and generate text feature vectors. Image feature extraction models are used to process image data from accident scenes and generate image feature vectors. The feature vectors of each modality are mapped to a unified feature space to generate a comprehensive feature representation.

3. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 2, characterized in that, The text feature extraction uses a pre-trained language model, and the calculation method is as follows: ； in For text feature vectors, This indicates the input text. Indicates a text encoder; The image feature extraction employs a convolutional neural network model, and the calculation method is as follows: ； in For image feature vectors, Indicates the input image. This indicates an image encoder.

4. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 1, characterized in that, The steps of generating high-quality fraud samples using the conditional generative adversarial network include: Construct a conditional generative adversarial network structure that includes a generator G and a discriminator D; The parameters of generator G and discriminator D are optimized through an adversarial training process that minimizes generator loss and maximizes discriminator loss. The generated fraudulent samples are evaluated and screened for quality, and high-quality samples are retained for subsequent model training.

5. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 4, characterized in that, The objective function of the adversarial training process is: ； in Indicates the generator The parameter minimization operation, Indicates the discriminator Parameter maximization operation Represents the value function. Indicates the expected value. This represents the log probability of the discriminator on the true sample. Represents the true sample distribution. This represents the complement of the logarithmic probability of the discriminator for the generated samples. Represents a random noise vector. Represents a condition vector. Represents a generator function; The sample quality assessment employs a method combining the output probability of discriminator D and a feature space distance metric, calculated as follows: ； in Indicates the generation of samples quality The weighting coefficients represent the output probability of the discriminator. This indicates that the discriminator evaluates the generated samples. The output probability, Indicates the generation of samples Compared with real fraud sample sets similarity, This represents the condition vector.

6. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 1, characterized in that, The steps for constructing the deep reconstruction network include: Construct a deep reconstruction network structure that includes an encoder and a decoder; The deep reconstruction network is trained by minimizing the reconstruction loss function, enabling it to accurately reconstruct the features of normal samples. For the sample to be detected, its reconstructed sample is calculated, and then a reconstruction error map is generated; Anomaly scores are calculated for samples based on reconstruction error maps, serving as an indicator of the likelihood of fraud.

7. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 6, characterized in that, The reconstruction loss function is defined as follows: ； in Indicates the losses incurred during reconstruction. This represents the reconstruction error of the sample space. Indicates sample features, Represents the reconstructed sample features. Denotes the square of the L2 norm. For feature extraction function, Indicates from original features The high-level feature representation extracted from it, Indicates from reconstruction features The high-level feature representation extracted from it, This represents the reconstruction error of the feature space. Weighting coefficients to balance the errors of the two parts; The abnormal score is calculated as follows: ； in Indicates sample Abnormal scores, Representing feature dimension, The first part represents the reconstruction error map. One element, Indicates the first The importance weights of each feature This indicates a summation operation across all feature dimensions.

8. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 1, characterized in that, The steps for constructing the difference-sensitive hash index include: The SMOTE variant algorithm is used to balance the dataset and increase the proportion of minority class samples. Based on the principle of locality-sensitive hashing, a set of hash functions is constructed to map the high-dimensional feature space to the low-dimensional hash code space. Enhance the difference sensitivity of locality-sensitive hash functions to improve their sensitivity to fraudulent features; A multi-level index structure is built based on hash encoding, which supports fast similarity retrieval.

9. The method for optimizing and retrieving vehicle insurance data for a digital management platform according to claim 8, characterized in that, The hash function is calculated as follows: ； in Representing the eigenvector One-dimensional hash code, It is a symbolic function. It is a random projection vector that follows a standard normal distribution. ; Points representing the projection vector and eigenvector; It is a random bias term that follows a uniform distribution. ; The enhanced difference sensitivity is achieved by weighting the projection vector according to the feature importance: ； in Represents element-level multiplication. Represents the original random projection vector. It is a feature importance vector. It is an adjustment factor. This represents the adjusted projection vector.

10. A vehicle insurance data optimization and retrieval system for a digital management platform, characterized in that: A method for optimizing and retrieving vehicle insurance data for a digital management platform as described in any one of claims 1-9, comprising: The multi-view feature extraction module is used to extract multimodal features from vehicle insurance data and map the extracted features to a unified feature space. The sample augmentation module is used to generate fraudulent samples based on the principle of conditional generative adversarial networks, balance the data distribution, and improve the model's generalization ability. The anomaly feature enhancement module is used to calculate the reconstruction error through a deep reconstruction network, amplify fraud feature signals, and generate anomaly scores. The high-efficiency indexing module is used to build differentially sensitive hash indexes based on anomaly features, enabling fast retrieval of fraudulent samples; The multi-view verification module is used to collect fraud evidence from multiple perspectives, assess consistency, calculate suspiciousness scores, and output the final detection results.

Images