Method, device and equipment for identifying vehicle insurance fraud based on new feature optimization algorithm

Abstract

The application discloses a car insurance fraud identification method, device and equipment based on a novel feature optimization algorithm, and the method comprises the following steps: acquiring a to-be-tested data set of car insurance claim fraud; performing classification processing on the to-be-tested data set based on a classification model to obtain a classification result; wherein the classification model is trained through a training set, and the training set is a data set that is balanced through a comprehensive sampling algorithm SMOTEEN and optimized through a feature selection algorithm EFS_BR proposed by the application; and the classification result is used for reflecting whether there is a car insurance claim fraud behavior. The application can achieve the effects of eliminating useless features, reducing the model calculation complexity and improving the model identification rate.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus and equipment for identifying car insurance fraud based on a novel feature optimization algorithm. Background Technology

[0002] The existing design principles of AI-based insurance fraud prevention models are as follows: based on a feature dataset of insurance fraud, a classification model is trained using machine learning or deep learning methods. This trained model is then used to predict whether insurance fraud will occur. Currently, researchers mostly use machine learning algorithms such as Bayesian networks, support vector machines, random forests, Gaussian anomaly detection, and clustering to achieve a risk identification rate of 80%.

[0003] While existing machine learning-based insurance fraud detection models can avoid the time-consuming and labor-intensive drawbacks of manual verification of fraudulent activities, they still have the following limitations:

[0004] 1. Supervised machine learning algorithms may suffer from classification bias when dealing with imbalanced insurance fraud sample data.

[0005] Classification bias refers to the uneven distribution of positive and negative samples in a dataset. In real-world scenarios, the number of negative samples indicating fraudulent behavior is far less than the number of normal car insurance claims. As a result, after training, while the model can efficiently identify normal samples, it struggles to correctly identify negative samples.

[0006] 2. Existing detection models based on feature engineering training do not further optimize features to improve model detection performance.

[0007] In real-world applications, researchers collect as many features as possible for training to improve detection performance. These feature sets may contain redundant features—features that have little impact on the model's classification performance or even negatively affect it. Furthermore, the higher the feature dimension, the greater the computational complexity.

[0008] 3. Traditional embedded feature selection algorithms ignore strongly correlated features and select weakly correlated features, which may lead to poor prediction performance of the model for future data. Summary of the Invention

[0009] This invention provides a method, apparatus, and device for identifying car insurance fraud based on a novel feature optimization algorithm, which solves the above-mentioned problems.

[0010] A method for identifying auto insurance fraud based on a novel feature optimization algorithm includes:

[0011] Obtain the dataset to be tested for auto insurance claims fraud;

[0012] The dataset to be tested is classified based on a classification model to obtain the classification result.

[0013] The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification results are used to reflect whether there is any fraudulent behavior in auto insurance claims.

[0014] In one embodiment of the present invention, during the optimization of the training set using the feature selection algorithm EFS_BR, the method further includes: obtaining a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; selecting s features from all features in the training set data matrix X using an embedded feature selection algorithm; determining the selected feature data matrix FS based on the selected s features; and determining the unselected feature data matrix UFS based on the unselected ds features.

[0015] In one embodiment of the present invention, during the optimization of the training set using the feature selection algorithm EFS_BR, the method further includes: determining the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; selecting k features with high correlation from the s selected features based on the correlation, and generating a first matrix about the k features; selecting t features with high correlation from the ds unselected features based on the correlation, and generating a second matrix about the t features; merging the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; training a base model based on the merged matrix W, wherein the base model employs a logistic regression classification algorithm; calculating the classification loss value of the base model; and determining the final output merged matrix W* based on the classification loss value as the optimized training set.

[0016] In one embodiment of the present invention, during the optimization of the training set using the feature selection algorithm EFS_BR, the method further includes: after determining the selected feature data matrix FS and the unselected feature data matrix UFS, entering an outer loop; wherein, the number of outer loop iterations is determined by the number of selected features s; selecting the top k features with high correlation from the selected feature data matrix FS to generate a first matrix, wherein k is equal to the current loop order of the outer loop; entering an inner loop, obtaining a merge matrix W based on the first matrix, training a base model based on the merge matrix W, wherein the base model uses a logistic regression classification algorithm; calculating the classification loss value of the base model based on the test set and its classification label data, and determining the merge matrix corresponding to the minimum classification loss value in the current loop; wherein, the feature categories in the test set are the same as the feature categories in the merge matrix; continuing the outer loop after the inner loop ends, returning the merge matrix W* corresponding to the minimum classification loss value after the outer loop ends, and obtaining the optimized training set based on the merge matrix W*.

[0017] In one embodiment of the present invention, obtaining the merged matrix based on the first matrix specifically includes: selecting the top t features with high correlation from the unselected feature data matrix to generate a second matrix, wherein t is equal to the current loop order of the inner loop; merging the first matrix and the second matrix to obtain the merged matrix.

[0018] In one embodiment of the present invention, determining the merge matrix corresponding to the minimum classification loss value in the current loop specifically includes: comparing the classification loss value of the base model with a preset initial classification loss value; if the calculated classification loss value is less than the initial classification loss value, then the calculated classification loss value is used as the initial classification loss value, and the merge matrix is ​​used as the feature data matrix corresponding to the minimum classification loss value.

[0019] In one embodiment of the present invention, the training set data is balanced using a comprehensive sampling algorithm, including: for samples in the training set, random sampling is performed from the category with fewer samples using the SMOTE algorithm, and the new samples obtained after sampling are added to the training set to obtain a pre-balanced training set; overlapping samples in the pre-balanced training set are removed using the ENN algorithm to obtain a final balanced training set.

[0020] In one embodiment of the present invention, training a classification model using a training set includes: training the classification model on a dataset optimized by the feature selection algorithm EFS_BR to obtain a classification model; wherein the classification model employs the logistic regression classification algorithm LR.

[0021] In one embodiment of the present invention, obtaining the test dataset for auto insurance claim fraud specifically includes: obtaining an initial dataset for auto insurance claim fraud detection; and selecting corresponding features from the initial dataset based on the optimization results of the EFS_BR algorithm to obtain the test dataset after feature selection.

[0022] A car insurance fraud detection device based on a novel feature optimization algorithm includes:

[0023] The data acquisition module is used to obtain the initial dataset for detecting fraud in auto insurance claims.

[0024] The model prediction module is used to classify the dataset to be tested based on the classification model to obtain the classification result; wherein, the classification model is trained using a training set, which is a dataset that has been balanced by a comprehensive sampling algorithm and optimized by the feature selection algorithm EFS_BR; the classification result is used to reflect whether there is any fraudulent behavior in car insurance claims.

[0025] A car insurance fraud detection device based on a novel feature optimization algorithm includes:

[0026] At least one processor; and,

[0027] The memory is communicatively connected to the at least one processor via a bus; wherein,

[0028] The memory stores instructions executable by the at least one processor, which are executed to perform:

[0029] Obtain the dataset to be tested for auto insurance claims fraud;

[0030] The dataset to be tested is classified based on a classification model to obtain the classification result.

[0031] The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification results are used to reflect whether there is any fraudulent behavior in auto insurance claims.

[0032] A non-volatile storage medium storing computer-executable instructions, which are executed by a processor to perform the following steps:

[0033] Obtain the dataset to be tested for auto insurance claims fraud;

[0034] The dataset to be tested is classified based on a classification model to obtain the classification result.

[0035] The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification results are used to reflect whether there is any fraudulent behavior in auto insurance claims.

[0036] This invention provides a method, apparatus, and device for identifying car insurance fraud based on a novel feature optimization algorithm, which has at least the following beneficial effects: by using the SMOTEEN comprehensive sampling algorithm to balance the collected training samples, the model's accurate identification rate of abnormal samples can be improved; by optimizing the collected feature samples, redundant features can be eliminated, the computational complexity of the model can be reduced, and the classification effect of the model can be further optimized. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0038] Figure 1 A schematic diagram illustrating the process of training an anti-fraud model using a training set, provided for an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the EFS_BR algorithm provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the steps of a car insurance fraud identification method based on a novel feature optimization algorithm provided in an embodiment of the present invention.

[0041] Figure 4 This is a schematic diagram illustrating a process for using a trained anti-fraud model to predict a dataset under test, as provided in an embodiment of the present invention.

[0042] Figure 5 A schematic diagram of a car insurance fraud identification device based on a novel feature optimization algorithm provided in an embodiment of the present invention;

[0043] Figure 6 This is a schematic diagram of a car insurance fraud identification device based on a novel feature optimization algorithm, provided as an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described clearly and completely below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0045] It should be noted that those skilled in the art will understand, explicitly and implicitly, that the embodiments described in this invention can be combined with other embodiments without conflict. Unless otherwise defined, the technical or scientific terms used in this invention should be understood in their ordinary sense by those skilled in the art. The terms "a," "an," "an," "the," etc., used in this invention do not indicate quantity limitation and can represent singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; the terms "first," "second," "third," etc., used in this invention are merely to distinguish similar objects and do not represent a specific ordering of objects.

[0046] This invention proposes a method, apparatus, and device for identifying auto insurance fraud based on a novel feature optimization algorithm. It utilizes machine learning algorithms to address the problem of imbalanced insurance fraud data in real-world scenarios; and employs an improved embedded feature selection algorithm, EFS_BR (Embedded Feature Selection based on Boosting Relevance), to remove useless features, reduce model computational complexity, and improve model recognition performance. A detailed explanation follows.

[0047] Before the model prediction module can identify the test data, the classification model within that module needs to be trained. The classification algorithm used in the model is Logistic Regression (LR). The LR algorithm is used to train the optimized feature sample data. The resulting classification model (anti-fraud model) is used to predict the category of the test data to determine whether car insurance claim fraud exists. When training the classification model, the optimized training set is required. For example... Figure 1 The diagram shown is a flowchart illustrating the process of training an anti-fraud model using a training set according to an embodiment of the present invention.

[0048] Before feature optimization, the training set needs to be balanced. Supervised machine learning algorithms often suffer from classification bias when dealing with imbalanced insurance fraud samples. In real-world scenarios, the number of negative samples indicating fraudulent activity is far less than the number of normal auto insurance claims. Without balancing the training set, the model, while efficiently identifying normal samples after training, struggles to correctly identify negative samples. Therefore, this invention addresses the first technical problem by using the SMOTEEN algorithm: how to use machine learning algorithms to handle imbalanced insurance fraud samples, thereby improving the accuracy of the classification model in identifying negative samples.

[0049] In one embodiment of the present invention, the training set data is balanced using a comprehensive sampling algorithm, including: for samples in the training set, random sampling is performed from the category with fewer samples using the SMOTE algorithm, and the new samples obtained after sampling are added to the training set to obtain a pre-balanced training set; overlapping samples in the pre-balanced training set are removed using the ENN algorithm to obtain a final balanced training set.

[0050] Specifically, the SMOTE algorithm and the ENN algorithm together form the SMOTEEN algorithm. The SMOTE algorithm is a random oversampling method commonly used in machine learning to handle imbalanced data. Its process involves randomly sampling from the class with fewer samples, then adding the new samples obtained through post-sampling processing to the dataset, thereby expanding the minority class sample and achieving the goal of balancing the data sample. In this invention, the data sample containing car insurance claim fraud is relatively small. The SMOTE algorithm is used to randomly sample from the data containing car insurance claim fraud, and the new samples obtained through post-sampling processing are added to the training set, thus expanding the data containing car insurance claim fraud.

[0051] However, this algorithm has a drawback: if the newly generated samples are distributed around the majority class samples, it will exacerbate the overlap between the two classes. Therefore, the Edited Nearest Neighbors (ENN) algorithm is needed to clean up overlapping samples. The ENN algorithm uses the idea of ​​clustering. If a sample belonging to the majority class has more than half of its K nearest neighbors that do not belong to the majority class, then this sample will be removed, and the remaining samples will be the final augmented samples. In this invention, if a data sample that does not involve car insurance claim fraud has more than half of its K nearest neighbors whose class labels are different, then this sample will be removed, and the remaining samples will form the final balanced dataset.

[0052] In real-world applications, researchers collect as many features as possible for training to improve detection and recognition performance. These feature sets may contain redundant features—features that have little or no impact on the model's classification performance or even reduce its effectiveness. Furthermore, the higher the feature dimension, the greater the computational complexity. Therefore, the second problem this invention aims to solve is how to use machine learning algorithms to further filter the collected feature samples, thereby eliminating redundant features, reducing the model's computational complexity, and ultimately further optimizing the model's classification performance.

[0053] Current feature optimization algorithms are divided into three categories:

[0054] (1) Filtering method. Features are ranked according to their divergence or correlation with the labels, and the top k features are selected. Commonly used metrics for measuring the correlation between features and labels include correlation coefficient, chi-square test, and mutual information.

[0055] (2) Wrapper method. This method selects the feature subset with the minimum classification loss based on the objective function. Commonly used wrapper feature selection algorithms include RFE and LVW.

[0056] (3) Embedded method. This method uses certain machine learning algorithms as base models for training, obtaining the weight coefficients for each feature. The weight coefficients represent the importance of the feature to the model. Features with weight coefficients greater than a set threshold are selected. Commonly used base models include penalty terms and tree models.

[0057] Their respective advantages and disadvantages are shown in Table 1:

[0058]

[0059] Table 1 Comparison of different feature-based optimization algorithms

[0060] Traditional embedded feature selection algorithms neglect strongly correlated features and select weakly correlated features, which may lead to poor prediction performance of the model for future data. Therefore, the third problem that this invention aims to solve is how to select features that are strongly correlated with the labels while using embedded feature selection algorithms.

[0061] This invention proposes a novel embedded feature selection algorithm, EFS_BR, which combines the advantages of the aforementioned feature optimization algorithms while avoiding their disadvantages, and optimizes the features of the training set.

[0062] In one embodiment of the present invention, the process of optimizing the training set using the feature selection algorithm EFS_BR includes: obtaining a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; selecting s features from all features in the training set data matrix X using an embedded feature selection algorithm; determining the selected feature data matrix FS based on the selected s features; and determining the unselected feature data matrix UFS based on the unselected ds features.

[0063] Specifically, the specific idea of ​​the feature selection algorithm EFS_BR proposed in this invention is as follows: Figure 2 Algorithm 1 is shown in the diagram. Now assume the training set has a total of d features and n samples, and the training set data is set as a matrix. The label data of the training set is a matrix If the number of selected features after using the traditional embedded feature selection algorithm is *s*, then the number of unselected features is *ds*. Therefore, the selected feature data matrix is... Unselected feature data matrix

[0064] In one embodiment of the present invention, the process of optimizing the training set using the feature selection algorithm EFS_BR includes: determining the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; selecting k features with high correlation from the s selected features based on the correlation, and generating a first matrix about the k features; selecting t features with high correlation from the ds unselected features based on the correlation, and generating a second matrix about the t features; merging the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; training a base model based on the merged matrix W, wherein the base model uses a logistic regression classification algorithm; calculating the classification loss value of the base model, and determining the final output merged matrix W* based on the classification loss value, which serves as the optimized training set.

[0065] Furthermore, the process of optimizing the training set using the feature selection algorithm EFS_BR includes: after determining the selected feature data matrix FS and the unselected feature data matrix UFS, entering the outer loop; wherein, the number of outer loop iterations is s based on the number of selected features s; selecting the top k features with high correlation from the selected feature data matrix FS to generate the first matrix, where k is equal to the current loop order of the outer loop; entering the inner loop, obtaining the merge matrix W based on the first matrix, training the base model based on the merge matrix W, and the base model adopting the logistic regression classification algorithm; calculating the classification loss value of the base model based on the test set and its classification label data, and determining the merge matrix corresponding to the minimum classification loss value in the current loop; wherein, the feature categories in the test set are the same as the feature categories in the merge matrix; after the inner loop ends, continuing the outer loop, and after the outer loop ends, returning the merge matrix W* corresponding to the minimum classification loss value, and obtaining the optimized training set based on the merge matrix W*.

[0066] Further, the merged matrix is ​​obtained based on the first matrix, specifically including: selecting the top t features with high correlation from the unselected feature data matrix to generate a second matrix, where t is equal to the current loop order of the inner loop; merging the first matrix and the second matrix to obtain the merged matrix.

[0067] Specifically, the final feature training data matrix W* obtained after EFS_BR processing corresponds to the test set X composed of the features. test The label data for the test set is y. test .

[0068] As shown in the embodiment of Algorithm 1, this invention uses a chi-square score-based filtering feature selection algorithm to select a subset of features with high relevance from the feature set of FS to obtain FS'. Simultaneously, a chi-square score-based filtering feature selection algorithm is used to select a subset of features with high relevance from the feature set of UFS to obtain UFS'. FS' and UFS' are then merged through a finite number of iterations, that is, the features with high relevance from the unselected UFS are used to replace the subset of features with low relevance from the selected FS, thereby improving the relevance between features and labels in FS.

[0069] Further, determining the merge matrix corresponding to the minimum classification loss value in the current loop specifically includes: comparing the classification loss value of the base model with the preset initial classification loss value; if the calculated classification loss value is less than the initial classification loss value, then the calculated classification loss value is used as the initial classification loss value, and the merge matrix is ​​used as the feature data matrix corresponding to the minimum classification loss value.

[0070] Specifically, after merging the two matrices, the base model using the logistic regression algorithm is trained based on the merged matrix W. The classification loss value of the base model is calculated based on the matrix of the test set with the same features as the merged matrix and the label matrix of the test set. The classification loss value is compared with the initial classification loss value. If it is less than the initial classification loss value, the classification loss value is assigned as the initial classification loss value, and the merged matrix is ​​used as the feature data matrix corresponding to the minimum classification loss value.

[0071] After the loop in the EFS_BR algorithm is completed, the feature data matrix W* corresponding to the minimum classification loss value is output. The data corresponding to this matrix is ​​used as the labeled training set after feature optimization.

[0072] By replacing the subset of features with low correlation in the selected FS feature set with features with high correlation in the unselected UFS feature set, the correlation between features in the training set and labels is improved. Based on the traditional embedded feature selection algorithm, features with strong correlation with labels are selected to obtain a new embedded feature optimization algorithm, EFS_BR, which can further improve the model recognition performance.

[0073] In one embodiment of the present invention, training a classification model using a training set includes: training the classification model on a dataset optimized by the feature selection algorithm EFS_BR to obtain a classification model; wherein the classification model employs the logistic regression classification algorithm LR.

[0074] Specifically, the classification algorithm uses Logistic Regression (LR). The LR algorithm is used to train the optimized feature sample data, and the resulting classification model is used to predict the category of the test data.

[0075] Figure 3 A schematic diagram illustrating the steps of a car insurance fraud identification method based on a novel feature optimization algorithm provided in this embodiment of the invention may include the following steps:

[0076] S310: Obtain the test dataset for auto insurance claims fraud.

[0077] In one embodiment of the present invention, obtaining a test dataset for auto insurance claim fraud specifically includes: obtaining an initial dataset for auto insurance claim fraud detection; and selecting corresponding features from the initial dataset based on the optimization results of the EFS_BR algorithm to obtain a test dataset after feature selection.

[0078] Specifically, such as Figure 4 As shown, the data acquisition module collects features for detecting auto insurance claim fraud, constructing an initial dataset for auto insurance claim fraud. The features used in this invention mainly include vehicle owner personal information, vehicle information, accident information, and claim information. Personal information includes the driver's age, gender, marital status, personal income, and education level. Vehicle information includes safety rating, vehicle weight, color, and age. Accident information includes the accident location and whether there were witnesses at the accident scene. Claim information includes the number of past claims, reporting channels, and compensation amounts.

[0079] After obtaining the initial dataset, feature filtering is performed on all features acquired in the data acquisition module based on the feature optimization results of the EFS_BR algorithm obtained during model training. This aims to eliminate redundant features, reduce model computational complexity, and further optimize the model's classification performance. The data obtained after feature filtering is the test dataset.

[0080] S320: The classification model is used to classify the dataset to be tested to obtain the classification result; the classification model is trained using a training set, which is a dataset that has been balanced by a comprehensive sampling algorithm and optimized by the feature selection algorithm EFS_BR; the classification result is used to reflect whether there is any fraudulent behavior in car insurance claims.

[0081] Specifically, the model prediction module classifies the test dataset to determine whether there is any car insurance claim fraud. The classification algorithm uses Logistic Regression (LR). The LR algorithm is used to train the data with optimized features, and the resulting classification model is used to predict the category of the sample data.

[0082] In one embodiment of the present invention, to verify the actual effect of the invention, a publicly available dataset is used for verification. This dataset contains 17,995 labeled samples, with a ratio of fraudulently labeled samples to normal samples of approximately 1:5, and a training set to test set split ratio of 8:2. In imbalanced classification tasks, accuracy cannot be used as an evaluation metric for model performance; therefore, this invention uses precision to evaluate the model's effectiveness.

[0083] The optimized model results are shown in Table 2:

[0084]

[0085] Table 2 Model Classification Results

[0086] Table 2 shows that when only LR is used for classification, the uneven distribution of the two classes of samples leads to poor model recognition performance, with an accuracy of only 0.846. After using the SMOTE algorithm to equalize the samples, the recognition accuracy improved by 6.1%, reaching 90.7%. Further, by removing overlapping samples and reducing noise, the accuracy improved by another 3.2%. Using the traditional embedded feature selection algorithm, the model's recognition accuracy improved by another 0.1%, reaching 94%. Using the feature selection algorithm EFS_BR proposed in this invention, the model's recognition accuracy improved by another 0.3%, reaching 94.3%. The comparison of these models verifies that the anti-fraud model proposed in this invention can effectively improve its recognition accuracy through imbalanced data processing and feature optimization. Furthermore, the improved feature selection algorithm EFS_BR outperforms the traditional embedded feature selection algorithm, and the final model performance surpasses most current research results.

[0087] The above describes a method for identifying car insurance fraud based on a novel feature optimization algorithm, as provided in this invention. Based on the same inventive concept, this invention also provides a corresponding device for identifying car insurance fraud based on a novel feature optimization algorithm, such as... Figure 5 As shown.

[0088] Data acquisition module 502 is used to acquire the dataset to be tested for car insurance claim fraud detection;

[0089] The model prediction module 504 is used to classify the dataset to be tested based on the classification model and obtain the classification result. The classification model is trained using a training set, which is a dataset that has been balanced by a comprehensive sampling algorithm and optimized by the feature selection algorithm EFS_BR. The classification result is used to reflect whether there is any fraudulent behavior in car insurance claims.

[0090] This invention also provides a corresponding vehicle insurance fraud identification device based on a novel feature optimization algorithm, such as... Figure 6 As shown, it includes:

[0091] The system includes at least one processor 602, a communication interface 604, a memory 606, and a communication bus 608; wherein the processor 602, the communication interface 604, and the memory 606 communicate with each other via the communication bus 608; the processor 602 can call logical instructions stored in the memory 606 to cause at least one processor 602 to execute:

[0092] Obtain a test dataset of car insurance claim fraud; classify the test dataset based on a classification model to obtain classification results; the classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR; the classification results are used to reflect whether car insurance claim fraud exists.

[0093] Based on the same idea, some embodiments of the present invention also provide media corresponding to the above methods.

[0094] Some embodiments of the present invention provide a storage medium storing computer-executable instructions, which are executed by a processor to perform the following steps:

[0095] Obtain a test dataset of car insurance claim fraud; classify the test dataset based on a classification model to obtain classification results; the classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR; the classification results are used to reflect whether car insurance claim fraud exists.

[0096] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are relatively simple in description because they are fundamentally similar to the method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0097] The devices, media, and methods provided in the embodiments of the present invention are one-to-one correspondences. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0098] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process method or product that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process method or product. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process method or product that includes that element.

[0099] The above are merely embodiments of the present invention and are not intended to limit the invention. Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for identifying car insurance fraud based on a novel feature optimization algorithm, characterized in that, include: Obtain the dataset to be tested for auto insurance claims fraud; The dataset to be tested is classified based on a classification model to obtain the classification result. The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification result is used to reflect whether there is any car insurance claim fraud. The features used in the training set include the vehicle owner's personal information, vehicle information, accident information, and claim information. The training set data is balanced using a comprehensive sampling algorithm, which includes: for samples in the training set, random sampling is performed from the category with fewer samples using the SMOTE algorithm, and the new samples obtained after sampling are added to the training set to obtain a preliminary balanced training set; overlapping samples in the preliminary balanced training set are removed using the ENN algorithm to obtain the final balanced training set. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: obtaining a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; selecting s features from all features in the training set data matrix X using an embedded feature selection algorithm; determining the selected feature data matrix FS based on the selected s features; and determining the unselected feature data matrix UFS based on the unselected ds features. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: determining the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; selecting k features with high correlation from the s selected features based on the correlation, and generating a first matrix about the k features; selecting t features with high correlation from the ds unselected features based on the correlation, and generating a second matrix about the t features; merging the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; training a base model based on the merged matrix W, wherein the base model uses a logistic regression classification algorithm; calculating the classification loss value of the base model, and determining the final output merged matrix W* based on the classification loss value, which serves as the optimized training set.

2. The method according to claim 1, characterized in that, The method further includes the following steps in optimizing the training set using the feature selection algorithm EFS_BR: After determining the selected feature data matrix FS and the unselected feature data matrix UFS, the outer loop is entered; the number of iterations in the outer loop is determined by the number of selected features s. The first matrix is ​​generated by selecting the top k features with high correlation from the selected feature data matrix FS, where k is equal to the current loop order of the outer loop; Enter the inner loop, obtain the merge matrix W based on the first matrix, and train the base model based on the merge matrix W. The base model adopts the logistic regression classification algorithm. The classification loss value of the base model is calculated based on the test set and its classification label data, and the merge matrix corresponding to the minimum classification loss value in the current loop is determined; wherein, the feature categories in the test set are the same as the feature categories in the merge matrix; After the inner loop finishes, the outer loop continues. After the outer loop finishes, the merge matrix W* corresponding to the minimum classification loss value is returned. The optimized training set is obtained based on the merge matrix W*.

3. The method according to claim 2, characterized in that, The step of obtaining the merged matrix based on the first matrix specifically includes: The first t highly correlated features are selected from the unselected feature data matrix to generate a second matrix, where t is equal to the current loop order of the inner loop; The first matrix and the second matrix are merged to obtain the merged matrix.

4. The method according to claim 2, characterized in that, Determining the merge matrix corresponding to the minimum classification loss value in the current loop specifically includes: The classification loss value of the base model is compared with the preset initial classification loss value; If the calculated classification loss value is less than the initial classification loss value, then the calculated classification loss value is used as the initial classification loss value, and the merge matrix is ​​used as the feature data matrix corresponding to the minimum classification loss value.

5. The method according to claim 1, characterized in that, Training the classification model using the training set includes: The classification model is trained using the dataset optimized by the feature selection algorithm EFS_BR to obtain the classification model; The classification model uses the logistic regression (LR) algorithm.

6. The method according to claim 1, characterized in that, The dataset for obtaining the test case of auto insurance claim fraud specifically includes: Obtain the initial dataset for detecting auto insurance claims fraud; Based on the optimization results of the EFS_BR algorithm, corresponding features are selected from the initial dataset to obtain the test dataset after feature selection.

7. A vehicle fraud detection device based on a novel feature optimization algorithm, characterized in that, include: The data acquisition module is used to acquire the dataset to be tested for detecting fraud in auto insurance claims. The model prediction module is used to classify the dataset to be tested based on a classification model to obtain a classification result. The classification model is trained using a training set, which is a dataset balanced by a comprehensive sampling algorithm and optimized by the feature selection algorithm EFS_BR. The classification result is used to reflect whether there is any car insurance claim fraud. The features used in the training set include vehicle owner personal information, vehicle information, accident information, and claim information. The device is also used to: balance the training set data using a comprehensive sampling algorithm, including: for samples in the training set, randomly sampling from the category with fewer samples using the SMOTE algorithm, and then adding the new samples obtained after sampling to the training set to obtain a pre-balanced training set; and removing overlapping samples from the pre-balanced training set using the ENN algorithm to obtain a final balanced training set. The apparatus is further configured to: during the optimization of the training set using the feature selection algorithm EFS_BR, obtain a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; select s features from all features in the training set data matrix X using an embedded feature selection algorithm; determine the selected feature data matrix FS based on the selected s features; and determine the unselected feature data matrix UFS based on the unselected ds features. The apparatus is further configured to: during the optimization of the training set using the feature selection algorithm EFS_BR, determine the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; based on the correlation, select k features with high correlation from the s selected features and generate a first matrix about the k features; based on the correlation, select t features with high correlation from the ds unselected features and generate a second matrix about the t features; merge the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; train a base model based on the merged matrix W, wherein the base model employs a logistic regression classification algorithm; calculate the classification loss value of the base model, and determine the final output merged matrix W* based on the classification loss value as the optimized training set.

8. A car insurance fraud detection device based on a novel feature optimization algorithm, characterized in that, include: At least one processor; as well as, The memory is communicatively connected to the at least one processor via a bus; wherein, The memory stores instructions executable by the at least one processor, which are executed to perform: Obtain the dataset to be tested for auto insurance claims fraud; The dataset to be tested is classified based on a classification model to obtain the classification result. The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification result is used to reflect whether there is any car insurance claim fraud. The features used in the training set include vehicle owner personal information, vehicle information, accident information, and claim information. The training set data is balanced using a comprehensive sampling algorithm, which includes: for samples in the training set, random sampling is performed from the category with fewer samples using the SMOTE algorithm, and the new samples obtained after sampling are added to the training set to obtain a preliminary balanced training set; overlapping samples in the preliminary balanced training set are removed using the ENN algorithm to obtain the final balanced training set. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: obtaining a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; selecting s features from all features in the training set data matrix X using an embedded feature selection algorithm; determining the selected feature data matrix FS based on the selected s features; and determining the unselected feature data matrix UFS based on the unselected ds features. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: determining the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; selecting k features with high correlation from the s selected features based on the correlation, and generating a first matrix about the k features; selecting t features with high correlation from the ds unselected features based on the correlation, and generating a second matrix about the t features; merging the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; training a base model based on the merged matrix W, wherein the base model uses a logistic regression classification algorithm; calculating the classification loss value of the base model, and determining the final output merged matrix W* based on the classification loss value, which serves as the optimized training set.

9. A non-volatile storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are executed by the processor to perform the following steps: Obtain the dataset to be tested for auto insurance claims fraud; The dataset to be tested is classified based on a classification model to obtain the classification result. The classification model is trained using a training set, which is a dataset that has been balanced using a comprehensive sampling algorithm and optimized using the feature selection algorithm EFS_BR. The classification result is used to reflect whether there is any car insurance claim fraud. The features used in the training set include vehicle owner personal information, vehicle information, accident information, and claim information. The training set data is balanced using a comprehensive sampling algorithm, which includes: for samples in the training set, random sampling is performed from the category with fewer samples using the SMOTE algorithm, and the new samples obtained after sampling are added to the training set to obtain a preliminary balanced training set; overlapping samples in the preliminary balanced training set are removed using the ENN algorithm to obtain the final balanced training set. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: obtaining a balanced training set data matrix X, wherein the training set data matrix X includes n samples and d features; selecting s features from all features in the training set data matrix X using an embedded feature selection algorithm; determining the selected feature data matrix FS based on the selected s features; and determining the unselected feature data matrix UFS based on the unselected ds features. In the process of optimizing the training set using the feature selection algorithm EFS_BR, the method further includes: determining the correlation between each feature and label in the selected feature data matrix FS and the unselected feature data matrix UFS using a filtering feature selection algorithm; selecting k features with high correlation from the s selected features based on the correlation, and generating a first matrix about the k features; selecting t features with high correlation from the ds unselected features based on the correlation, and generating a second matrix about the t features; merging the first matrix and the second matrix to obtain a merged matrix W, thereby removing features in FS whose correlation does not reach the correlation threshold; training a base model based on the merged matrix W, wherein the base model uses a logistic regression classification algorithm; calculating the classification loss value of the base model, and determining the final output merged matrix W* based on the classification loss value, which serves as the optimized training set.

Images