Insurance fraud detection method, related device, equipment and storage medium

Abstract

The invention discloses an insurance fraud detection method, a related device, equipment and a storage medium, and the method comprises the steps: carrying out the feature extraction based on to-be-detected insurance data, and obtaining a first scale feature, a second scale feature, and a first fusion feature of the first scale feature and the second scale feature; based on the first fusion feature, predicting to obtain a first weight representation and a second weight representation; obtaining a second fusion feature based on a weighted feature of the first scale feature by the first weight representation and a weighted feature of the second scale feature by the second weight representation; performing classification prediction based on the second fusion feature to obtain a detection result of the to-be-detected insurance data; wherein the detection result at least comprises whether the to-be-detected insurance data is suspected of fraud. According to the scheme, the insurance fraud detection efficiency and precision can be improved, and the insurance fraud detection cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an insurance fraud detection method and related devices, equipment and storage media. Background Art

[0002] Insurance fraud detection is crucial to preventing insurance institutions from suffering financial losses and maintaining the continued operation of the insurance system.

[0003] At present, the process of insurance fraud detection still relies on manual identification, which is not only inefficient but also costly. In addition, thanks to the rapid development of machine learning, machine learning technology is also being used to detect insurance fraud. However, these existing methods are highly dependent on insurance data feature selection or manual design of insurance data features. The former may lead to information loss, and the latter cannot fully capture the deep structure of insurance data, both of which greatly restrict the accuracy of insurance fraud detection. In view of this, how to improve the efficiency and accuracy of insurance fraud detection and reduce the cost of insurance fraud detection has become an urgent problem to be solved. Summary of the invention

[0004] The main technical problem solved by this application is to provide an insurance fraud detection method and related devices, equipment and storage media, which can improve the efficiency and accuracy of insurance fraud detection and reduce the cost of insurance fraud detection.

[0005] In order to solve the above technical problems, the first aspect of the present application provides an insurance fraud detection method, including: performing feature extraction based on the insurance data to be tested to obtain a first scale feature, a second scale feature, and a first fusion feature of the first scale feature and the second scale feature; based on the first fusion feature, predicting a first weight representation and a second weight representation; obtaining a second fusion feature based on the weighted feature of the first scale feature of the first weight representation and the weighted feature of the second scale feature of the second weight representation; performing classification prediction based on the second fusion feature to obtain a detection result of the insurance data to be tested; wherein the detection result at least includes whether the insurance data to be tested is suspected of fraud.

[0006] In order to solve the above technical problems, the second aspect of the present application provides an insurance fraud detection device, including: a feature extraction module, a weight prediction module, a weighted fusion module and a classification prediction module, the feature extraction module is used to perform feature extraction based on the insurance data to be tested, and obtain a first scale feature, a second scale feature and a first fusion feature of the first scale feature and the second scale feature; the weight prediction module is used to perform prediction based on the first fusion feature to obtain a first weight and a second weight; the weighted fusion module is used to perform classification prediction based on the first weight of the first scale feature and the second weighted feature of the second scale feature to obtain a second fusion feature; the classification prediction module is used to perform classification prediction based on the second fusion feature to obtain a detection result of the insurance data to be tested; wherein the detection result at least includes whether the insurance data to be tested is suspected of fraud.

[0007] In order to solve the above technical problems, the third aspect of the present application provides an electronic device, which at least includes a memory and a processor coupled to each other, the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the insurance fraud detection method in the above first aspect.

[0008] In order to solve the above technical problems, the fourth aspect of the present application provides a computer-readable storage medium storing program instructions that can be executed by a processor, and the program instructions are used to implement the insurance fraud detection method of the first aspect.

[0009] The above scheme extracts features based on the insurance data to be tested, obtains the first scale feature, the second scale feature and the first fusion feature of the first scale feature and the second scale feature, and predicts the first weight representation and the second weight representation based on the first fusion feature, and then obtains the second fusion feature based on the weighted feature of the first scale feature of the first weight representation and the weighted feature of the second scale feature of the second weight representation, and then performs classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested, and the detection result at least includes whether the insurance data to be tested is suspected of fraud. Therefore, on the one hand, since there is no need to manually identify the insurance data to be tested, the cost of insurance fraud detection can be reduced and the efficiency of insurance fraud detection can be improved. On the other hand, by performing weight prediction on the first fusion feature that integrates multiple scale features, the prediction accuracy of the reference degree of different scale features can be improved. Then, different scales are weighted respectively, which can not only accurately focus on each feature, but also refer to different scale features at the same time, avoid information loss as much as possible, and comprehensively capture the deep structure of the insurance data to be tested, which helps to improve the accuracy of insurance fraud detection. Therefore, the efficiency and accuracy of insurance fraud detection can be improved, and the cost of insurance fraud detection can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1It is a flowchart of an embodiment of the insurance fraud detection method of the present application; Figure 2a This is a schematic diagram of a process of performing initial encoding of insurance data to be tested in an embodiment; Figure 2b is a schematic diagram of a process of extracting features from coding features according to an embodiment; Figure 2c is a schematic diagram of a process of fusing features of different scales in one embodiment; Figure 2d is a schematic diagram of a process of weight prediction embodiment; Figure 2e is a process schematic diagram of an embodiment of weighted fusion; Figure 3 It is a schematic diagram of the framework of an embodiment of the insurance fraud detection device of the present application; Figure 4 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application; Figure 5 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0011] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0012] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0013] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the fragment " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0014] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the insurance fraud detection method of the present application. Specifically, it may include the following steps: Step S11: extracting features based on the insurance data to be tested to obtain a first scale feature, a second scale feature, and a first fusion feature of the first scale feature and the second scale feature.

[0015] In one implementation scenario, the initial data features of the insurance data to be tested can be extracted first, and then a first extraction can be performed based on the initial data features to obtain a first scale feature, and a second extraction can be performed based on the initial data features to obtain a second scale feature, and the size of the convolution kernel used in the first extraction is smaller than the convolution kernel used in the second extraction. On this basis, multi-scale features can be obtained by fusion based on the first scale feature and the second scale feature, and then at least convolution processing can be performed based on the multi-scale feature to obtain a first fused feature. The above method extracts features of the initial data features of the insurance data to be tested by a feature extraction method including convolution kernels of different sizes to obtain features of different scales, and then multi-scale fusion and convolution processing are performed to deeply fuse features of different scales, which can promote feature fusion of multi-scale features as much as possible, and help capture the complex correlation between the features of the insurance data to be tested.

[0016] In a specific implementation scenario, the insurance data to be tested may include but not be limited to: basic insurance information, basic information of the insured, basic information of the insured's claim accident, and other dimensions. The coverage dimensions of the insurance data to be tested are not limited here. For example, the basic information of the insurance project may include but is not limited to: insurance number, age, time of becoming a customer (in months), insurance liability limit, etc.; the basic information of the insured may include but is not limited to: the insured's zip code, gender, occupation, hobbies, insured relationship, capital gains, capital losses, etc.; the basic information of the insured's claim accident may include but is not limited to: including the date of the accident, the year of car purchase, the type of accident, the type of collision, the number of vehicles involved, whether there is property damage, the degree of physical injury, witnesses, whether there is a record report, the overall claim amount, the injury claim amount, the property claim amount, the car claim amount, the car brand and model, etc.

[0017] In a specific implementation scenario, before extracting features from the insurance data to be tested, the insurance data to be tested may be preprocessed, which may include but is not limited to: missing value processing, outlier processing, elimination of invalid values, etc. The specific operations of the preprocessing are not limited here.

[0018] In a specific implementation scenario, please refer to Figure 2a , Figure 2a FIG. 1 is a schematic diagram of an embodiment of the process of initial encoding of the insurance data to be tested. Figure 2aAs shown, the discrete values ​​of the insurance data to be tested (such as age, gender, etc.) can be one-hot encoded to obtain the first encoded feature. It should be noted that one-hot encoding converts each discrete value into a set of binary features, which effectively reduces the weight of a single feature in the model, reduces the impact of abnormal data on the model, and expands the data features to a certain extent. This processing method not only enhances the robustness of the model, but also improves the ability to handle discrete features, laying the foundation for subsequent data analysis and machine learning modeling. Suppose there is an insurance classification variable X, which has n possible values , the one-hot encoded vector is represented as V, with a length of n. For each value , the encoded vector The i-th dimension of is 1, and the other dimensions are 0. Specifically, the mathematical formula of one-hot encoding can be expressed as: Among them, if ,but ,if ,but . At the same time, the continuous values ​​in the tested insurance data can be normalized and standardized to obtain the second coding feature. In insurance data analysis and risk management, normalization and standardization are important steps to improve model performance. The dimensions and orders of magnitude of different risk factors may vary greatly. Through normalization or standardization, these factors can be compared on the same scale, thereby more accurately assessing the overall risk. In addition, for insurance product pricing models involving multiple indicators, normalization and standardization can also help improve the accuracy and stability of the model. Specifically, normalization is the change of data to a fixed interval (range). Usually, this interval is a decimal between [0,1] or (-1,1). It is mainly proposed for the convenience of data processing. Mapping the data to the range of 0 to 1 for processing is more convenient and fast, and it is easy to eliminate the dimension and incorporate the data of each indicator into the comprehensive evaluation. For example, the formula can be used Normalization operation is performed to facilitate the subsequent gradient descent and activation function processing of data. Represents any numeric data. Represents the minimum value in a series of numerical data. Represents the maximum value in a series of numerical data. After standardization, the insurance data is distributed around 0, and the functions sigmoid, Tanh, Softmax, etc. are also distributed around 0; the value in each feature becomes 0 on average (the value of each feature is subtracted from the average value of the feature in the original insurance data). , the standard deviation becomes 1, the formula Indicates a standardized operation. Figure 2b , Figure 2bThis is a schematic diagram of an embodiment of extracting features from coding features. Figure 2b As shown, after encoding to obtain the first encoding feature and the second encoding feature, feature extraction can be performed based on the first encoding feature to obtain the first initial feature, and feature extraction can be performed based on the second encoding feature to obtain the second initial feature, so that the first initial feature and the second initial feature can be fused to obtain the initial data feature of the insurance data to be tested. Exemplarily, the fully connected neural network has shown strong capabilities and wide application potential in processing insurance data feature extraction. Its high connectivity and nonlinear processing capabilities ensure efficient processing of complex data, while its flexibility and optimizability enable the network to be adjusted according to specific tasks. These advantages provide strong technical support for insurance data analysis and help improve the accuracy of risk management and prediction. The fully connected neural network automatically adjusts feature weights through the training process, strengthens important insurance data features, and weakens unimportant insurance data features. This weight optimization enables the network to autonomously find the most discriminating insurance features when facing large-scale insurance data sets. The fully connected layer can fuse different types of input insurance data features into a unified insurance data feature space. This is crucial for the basic information of comprehensive insurance projects, insured information, and claim accident information. The fully connected neural network can process high-dimensional data containing a large number of features. Each neuron is connected to all neurons in the previous layer, ensuring that even subtle insurance data features are not lost. This structure enables the network to capture subtle patterns in the data, which is particularly important for complex insurance data. Nonlinear transformations provide better fitting and generalization capabilities for complex relationships and patterns commonly found in insurance data. Therefore, for the first encoded feature and the second encoded feature, a fully connected neural network can be used for feature extraction to obtain the first initial feature and the second initial feature. For example, a three-layer fully connected neural network can be used, with the number of neurons in each fully connected layer being 16, 32, and 16, respectively, and the activation function can be set to ReLu (i.e., linear rectifier activation). On this basis, the first initial feature and the second initial feature can be concatenated to achieve feature fusion and obtain the initial data feature. In the above method, discrete values ​​in the insurance data to be tested are one-hot encoded to obtain a first encoding feature, and continuous values ​​in the insurance data to be tested are normalized and standardized to obtain a second encoding feature, and feature extraction is then performed based on the first encoding feature to obtain a first initial feature, and feature extraction is performed based on the second encoding feature to obtain a second initial feature, thereby fusing the first initial feature and the second initial feature to obtain the initial data feature of the insurance data to be tested, and then initial encoding and feature extraction can be performed on discrete values ​​and continuous values ​​in the insurance data to be tested, respectively, which helps to capture the deep structure of the insurance data to be tested as completely and fully as possible.

[0019] In a specific implementation scenario, as a possible example, the convolution kernel used in the first extraction can be a 1*1 one-dimensional convolution kernel (the number of channels can be 1), and the convolution kernel used in the second extraction can be a 1*2 one-dimensional convolution kernel, which has a larger receptive field (it can be filled with 0 first and then perform a convolution operation to obtain the same dimension as the first scale feature). The specific sizes of the convolution kernels used in the first extraction and the second extraction are not limited here.

[0020] In a specific implementation scenario, as a possible example, the first extraction may include sequentially executing one-dimensional convolution (such as the aforementioned 1*1 one-dimensional convolution), batch normalization, pooling, and linear rectification activation. It should be noted that the batch normalization operation is used to normalize the number of columns of these input values ​​or tensors of the convolutional network and scale them to a suitable range, making it easier for the model to converge and allowing each layer to face input values ​​with the same feature distribution as much as possible, reducing the uncertainty caused by changes.

[0021] In a specific implementation scenario, as a possible example, the second extraction may include sequentially performing one-dimensional convolution (such as the aforementioned 1*2 one-dimensional convolution), batch normalization, pooling, linear rectification activation, one-dimensional convolution, batch normalization, pooling, and linear rectification activation.

[0022] In a specific implementation scenario, as a possible example, still taking the case where both the first extraction and the second extraction use one-dimensional convolution as an example, the first scale feature and the second scale feature can be obtained by, for example, one-dimensional feature fusion (1 Dimension Feature Fusion, 1DFF) to obtain multi-scale features. Of course, the above example is only a possible example of multi-scale fusion when both the first extraction and the second extraction use one-dimensional convolution, and other possible situations are not limited here.

[0023] In a specific implementation scenario, after obtaining the multi-scale features, at least one-dimensional convolution, batch normalization, pooling, linear rectification activation, one-dimensional convolution, linear rectification activation and full connection can be performed in sequence to make the first scale features and the second scale features in the multi-scale features as fully integrated as possible to obtain the first fused features. Of course, the above example is only a possible example of fully integrating the first scale features and the second scale features in the multi-scale features, and other possible methods are not limited thereto, and no examples are given here one by one.

[0024] In another implementation scenario, similar to the above implementation, the initial data features of the insurance data to be tested can be extracted first, and then the first extraction is performed based on the initial data features to obtain the first scale features, and the second extraction is performed based on the initial data features to obtain the second scale features, and the convolution kernel size used in the first extraction is smaller than the convolution kernel used in the second extraction, so as to fuse the multi-scale features based on the first scale features and the second scale features. Different from the above implementation, as another possible implementation, it can be detected whether the second number of times the first scale features and the second scale features perform multi-scale fusion reaches the second threshold. In response to the second number not reaching the second threshold, the multi-scale features can be selected as the new initial data features, and the first extraction based on the initial data features is returned to obtain the first scale features until the second number reaches the second threshold, and then the latest first scale features are selected as the final first scale features, and the latest second scale features are selected as the final second scale features. After that, the steps of fusing the first scale features and the second scale features in the above implementation and obtaining the first fused features, and performing at least convolution processing based on the multi-scale features to obtain the first fused features can be continued to obtain the first fused features of the final first scale features and the final second scale features. In the above manner, by detecting whether the second number of times of executing multi-scale fusion reaches the second threshold, multiple executions of multi-scale fusion are triggered, so that features of different scales can be integrated as fully as possible.

[0025] In a specific implementation scenario, the second threshold can be set according to actual application needs. For example, when the requirement for fully integrating different scale features is higher, the second threshold can be set as large as possible. On the contrary, when the requirement for fully integrating different scale features is relatively loose, the second threshold can be set appropriately smaller. Exemplarily, the second threshold can be set to 2, 3, 4, etc., and the specific value of the second threshold is not limited here.

[0026] In a specific implementation scenario, please refer to Figure 2c , Figure 2c This is a schematic diagram of the process of fusing features of different scales in one embodiment. For the convenience of description, the one-dimensional convolution in the first extraction and the second extraction can be denoted as C, the batch normalization can be denoted as B, the pooling can be denoted as P, and the linear rectification activation can be denoted as R. In addition, the full connection can be denoted as FC. Taking the second threshold as 2 as an example, the initial data features can be extracted (CBPR) to obtain the first scale feature : …… (1) In the above formula (1), X represents the initial data feature. At the same time, after the second extraction (CBPRCBPR), the second scale feature can be obtained : …… (2) On this basis, the first scale feature and the second scale feature can be obtained by multi-scale fusion (DFF). Since the second number of multi-scale fusion has not reached the second threshold 2 at this time, the multi-scale feature is selected as the new initial data feature, and the first extraction (CBPR) is performed on the initial data feature to obtain a new first scale feature, and the second extraction (CBPR) is performed on the initial data feature to obtain a new second scale feature. The new first scale feature and the new second scale feature are fused by multi-scale fusion (DFF) to obtain a multi-scale feature. Since the second number of multi-scale fusion reaches the second threshold 2 at this time, the latest first scale feature can be selected as the final first scale feature, and the latest second scale feature can be selected as the final second scale feature, and at least convolution processing (CBPRCR and FC) is performed on the latest multi-scale feature to obtain the first fusion feature. Of course, the above example is only a possible example of extracting the first scale feature and the second scale feature and the first fusion feature of the two when the second threshold is set to 2. When the second threshold is set to other values, it can be deduced by analogy, and no examples are given here.

[0027] Step S12: Based on the first fusion feature, a first weighted representation and a second weighted representation are predicted.

[0028] In one implementation scenario, as a possible implementation method, the first fused feature can be processed based on the attention mechanism to obtain a first weighted representation. It should be noted that the attention mechanism may include but is not limited to the self-attention mechanism, and the specific type of the attention mechanism is not limited here. In addition, for the specific process of processing the first fused feature based on the attention mechanism, please refer to the technical details of the attention mechanism, which will not be repeated here. The first weighted representation has the same size as the first fused feature. By way of example, taking the specific size of the first fused feature as C*H as an example, where C represents the number of channels of the first fused feature, and H represents the number of dimensions of the first fused feature, the specific size of the first weighted representation is C*H. After obtaining the first weighted representation, a second weighted representation that is negatively correlated with the first weighted representation can be obtained. By way of example, the second weighted representation can be obtained by subtracting each element in the first weighted representation from 1.

[0029] In another implementation scenario, different from the aforementioned implementation, global average pooling can also be performed based on the first fused feature to obtain a pooled dimensionality reduction feature. On this basis, weight prediction can be performed based on the first fused feature to obtain a local weight representation, and weight prediction can be performed based on the pooled dimensionality reduction feature to obtain a global weight representation, and the weight prediction includes at least one-dimensional convolution and batch normalization performed sequentially. On this basis, a fusion can be performed based on the global weight representation and the local weight representation to obtain a first weight representation, and the second weight representation is negatively correlated with the first weight representation. The above method performs weight prediction from both global and local channel dimensions, and obtains a weight representation by combining the two channel dimensions. It can maintain the global information of the global features while maintaining the original details of the local features, which helps to extract higher-level comprehensive features from the insurance data to be tested.

[0030] In a specific implementation scenario, please refer to Figure 2d , Figure 2d FIG. 1 is a schematic diagram of a weight prediction process according to an embodiment of the present invention. Figure 2d As shown, in the global channel dimension, the first fusion feature of size C*H is globally averaged pooled, and each channel can be compressed to 1*1, so a pooled dimension reduction feature of size C*1*1 can be obtained. On this basis, the pooled dimension reduction feature of size C*1*1 can be processed by one-dimensional convolution (the size can be 1*1) to obtain an intermediate feature of C / r*1*1, where r represents the size of the one-dimensional convolution. After batch normalization, linear rectification activation, one-dimensional convolution, and batch normalization, a global weight representation is obtained, and its size is the same as the first fusion feature, that is, C*H. For the convenience of description, the global weight representation G can be expressed as: …… (3) In the above formula (3), X represents the first fused feature, avg represents global average pooling, C represents one-dimensional convolution, B represents batch normalization, and R represents linear rectifier activation. In this way, the global features in the data set, that is, those significant features that dominate the overall distribution, can be captured. Similarly, in the local channel dimension, the first fused feature of size C*H can be formally flattened to C*1*1. On this basis, the pooled dimensionality reduction feature of size C*1*1 can be processed by one-dimensional convolution to obtain an intermediate feature of C / r*1*1, where r represents the size of the one-dimensional convolution. After batch normalization, linear rectifier activation, one-dimensional convolution, and batch normalization, a local weight representation is obtained, and its size is also the same as the first fused feature, that is, C*H. For the convenience of description, the local weight representation L can be expressed as: …… (4) In the above formula (4), X represents the first fused feature, C represents one-dimensional convolution, B represents batch normalization, and R represents linear rectification activation. In this way, some local and smaller features can be captured, which may be crucial to insurance decisions in specific cases even though they are not so significant in the global distribution.

[0031] In a specific implementation scenario, after obtaining the global weight representation and the local weight representation, element-by-element addition can be performed based on the global weight representation and the local weight representation to obtain a fused weight representation, and then each element in the fused weight representation is constrained to a preset weight range based on a preset activation function to obtain a first weight representation. Exemplarily, the preset activation function may include but is not limited to sigmoid, so as to constrain each element in the fused weight representation to a preset weight range of 0 to 1 to obtain a first weight representation. On this basis, each element in the first weight representation can be subtracted from 1 to obtain a second weight representation. In the above method, by adding the global weight representation and the local weight representation element-by-element to obtain a fused weight representation, and then constraining each element in the fused weight representation to a preset weight representation based on a preset activation function to obtain a first weight representation, the two dimensions of local channel and global channel can be integrated, which helps to improve the accuracy of the weight representation.

[0032] Step S13: Obtain a second fused feature based on the weighted feature of the first scale feature represented by the first weight and the weighted feature of the second scale feature represented by the second weight.

[0033] In one implementation scenario, as a possible example, the first scale feature can be weighted using the first weight representation, and the second scale feature can be weighted using the second weight representation, and the above two weighted features are fused (eg, added) to obtain a second fused feature.

[0034] In another implementation scenario, as another possible example, after obtaining the second fusion feature, it is also possible to detect whether the first number of weighted fusions of the first scale feature and the second scale feature reaches the first threshold. In response to the first number not reaching the first threshold, the second fusion feature can be selected as the new first fusion feature, and the prediction based on the first fusion feature is returned to obtain the first weight representation and the second weight representation step until the first number reaches the first threshold. At this point, the newly acquired second fusion feature can be used for subsequent classification predictions. The above method, by performing multiple weighted fusions of the first scale feature and the second scale feature, can reflect the inherent connection and characteristics of the insurance data to be tested as comprehensively as possible, which helps to accurately assess insurance risks.

[0035] In a specific implementation scenario, the first threshold can be set according to actual application needs. For example, when the requirements for weighted fusion are higher, the first threshold can be set as large as possible. Conversely, when the requirements for weighted fusion are relatively loose, the first threshold can be set appropriately smaller. Exemplarily, the first threshold can be set to 2, 3, 4, etc., and the specific value of the first threshold is not limited here.

[0036] In a specific implementation scenario, please refer to Figure 2e , Figure 2e FIG. 1 is a schematic diagram of a weighted fusion process according to an embodiment of the present invention. Figure 2e As shown, taking the first threshold value as 2 as an example, the first fusion feature After weight prediction (MC), the first weight representation and the second weight representation can be obtained. The first weight representation is for the first scale feature Weighted, the second weight represents the second scale feature After weighting, the two are fused to obtain the second fusion feature Z: …… (5) In the above formula (5), represents the first scale feature, represents the second scale feature, M represents the weight prediction MC, for details, please refer to Figure 2d And the related description will not be repeated here. represents the first weight representation, Represents the second weight representation. Based on this weighted fusion, we can understand the data from two different perspectives, which helps to better capture the associations and patterns in the data. Since the first number of weighted fusion is still 1 at this time, which does not reach the first threshold 2, the second fusion feature Z can be used as the new first fusion feature, and the above weight prediction is returned to obtain the second fusion feature used for classification prediction. : …… (6) Through the above two weighted fusions, the previously obtained information can be further integrated and optimized, reflecting the inherent connections and characteristics of the data more comprehensively, so as to more accurately assess insurance risks.

[0037] Step S14: Perform classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested.

[0038] In the disclosed embodiment, the detection result at least includes whether the insurance data to be tested is suspected of fraud. Of course, in addition to this, when the detection result includes that the insurance data to be tested is suspected of fraud, the detection result may also include a specific fraud type. The specific connotations of the detection result are not given one by one here.

[0039] In an implementation scenario, as described above, at least one of the first scale feature, the second scale feature, the first fusion feature, and the second fusion feature is obtained by a neural network model. As a possible implementation, a neural network-based classifier (such as a fully connected layer, etc.) can be further selected to classify and predict the second fusion feature to obtain the detection result of the insurance data to be tested.

[0040] In another implementation scenario, as mentioned above, at least one of the first scale feature, the second scale feature, the first fusion feature, and the second fusion feature is obtained by a neural network model. As another possible implementation, a classifier based on ensemble learning (such as XGBoost) can be selected to perform classification prediction on the second fusion feature to obtain the detection result of the insurance data to be tested. Taking XGBoost as a classifier as an example, the principle of XGBoost is to connect new decision trees in series under the Boosting framework, and improve the performance of the learner through multiple iterations of fitting residuals. Compared with the traditional gradient boosting tree algorithm, XGBoost uses a second-order Taylor expansion to approximate the loss function, and usually uses fewer decision trees to obtain higher accuracy. Assume a set of insurance sample data D: …… (7) In the above formula (7), is the insurance characteristic value, is the true value. XGBoost is an additive algorithm that adds the predicted probabilities of K trees to get the final prediction result : …… (8) In the above formula (8), is the final predicted value, represents the k-th decision tree function, and F represents the decision tree set. XGBoost is a model for learning K trees, so the objective function can be expressed as: …… (9) …… (10) In the above formulas (9) and (10), L is a differentiable loss function, is the true value and estimated values The difference, Regularization is added as a penalty term to constrain the growth of decision trees and avoid overfitting. In it, γ is a parameter representing complexity; T represents the number of leaf nodes; λ represents the penalty coefficient of leaf weight, which is usually a constant; Represents the weight value of the leaf node of the tree. The values ​​of γ and λ determine the complexity of the model and are usually set based on experience. The model adds a new decision tree function to the predicted value of the previous round to minimize the residual between the predicted value and the actual value. When the model has t trees, it is expressed as ,in, represents the predicted value of the previous round, then the objective function of the tth round becomes Therefore, the goal is to find the optimal solution that minimizes the objective function. And XGBoost rewrites the objective function, performs Taylor expansion, takes the first 3 terms, removes the higher-order infinitesimal terms, and finally transforms the objective function into: …… (11) In the above formula (10), and are the first-order derivative and second-order derivative of the loss function respectively. Since the residual of the previous round will not affect the optimization of the objective function, the obtained objective function can be approximately expressed as: …… (12) Convert the tree iteration model into the iteration form of leaf nodes to obtain the optimal leaf node value ,in, is all the first-order partial derivatives of the samples contained in leaf node j Similarly, Its second-order partial derivative Substituting the optimal value into the objective function, we finally get the minimum loss: …… (13) In general, XGBoost adds regularization to the standard function to reduce model complexity. It uses first-order and second-order derivatives to calculate losses more accurately. It also supports column sampling to avoid overfitting of the model while reducing the amount of calculation. Therefore, by combining the unique structure of neural networks and ensemble learning-based classifiers, efficient insurance feature extraction and accurate insurance fraud prediction can be achieved.

[0041] The above scheme extracts features based on the insurance data to be tested, obtains the first scale feature, the second scale feature and the first fusion feature of the first scale feature and the second scale feature, and predicts the first weight representation and the second weight representation based on the first fusion feature, and then obtains the second fusion feature based on the weighted feature of the first scale feature of the first weight representation and the weighted feature of the second scale feature of the second weight representation, and then performs classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested, and the detection result at least includes whether the insurance data to be tested is suspected of fraud. Therefore, on the one hand, since there is no need to manually identify the insurance data to be tested, the cost of insurance fraud detection can be reduced and the efficiency of insurance fraud detection can be improved. On the other hand, by performing weight prediction on the first fusion feature that integrates multiple scale features, the prediction accuracy of the reference degree of different scale features can be improved. Then, different scales are weighted respectively, which can not only accurately focus on each feature, but also refer to different scale features at the same time, avoid information loss as much as possible, and comprehensively capture the deep structure of the insurance data to be tested, which helps to improve the accuracy of insurance fraud detection. Therefore, the efficiency and accuracy of insurance fraud detection can be improved, and the cost of insurance fraud detection can be reduced.

[0042] See also Figure 3 , Figure 3 It is a schematic diagram of the framework of an embodiment of the insurance fraud detection device of the present application. The insurance fraud detection device 30 includes: a feature extraction module 31, a weight prediction module 32, a weighted fusion module 33 and a classification prediction module 34. The feature extraction module 31 is used to extract features based on the insurance data to be tested, and obtain the first scale feature, the second scale feature and the first fusion feature of the first scale feature and the second scale feature; the weight prediction module 32 is used to predict based on the first fusion feature to obtain the first weight and the second weight; the weighted fusion module 33 is used to obtain the second fusion feature based on the first weight of the first scale feature and the second weight of the second scale feature; the classification prediction module 34 is used to perform classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested; wherein the detection result at least includes whether the insurance data to be tested is suspected of fraud.

[0043] In the above scheme, the insurance fraud detection device 30 extracts features based on the insurance data to be tested, obtains the first scale feature, the second scale feature and the first fusion feature of the first scale feature and the second scale feature, and predicts the first weight representation and the second weight representation based on the first fusion feature, and then obtains the second fusion feature based on the weighted feature of the first weight representation on the first scale feature and the weighted feature of the second weight representation on the second scale feature, and then performs classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested, and the detection result at least includes whether the insurance data to be tested is suspected of fraud. Therefore, on the one hand, since there is no need to manually identify the insurance data to be tested, the cost of insurance fraud detection can be reduced and the efficiency of insurance fraud detection can be improved. On the other hand, by performing weight prediction on the first fusion feature that integrates multiple scale features, the prediction accuracy of the reference degree of different scale features can be improved. Then, different scales are weighted respectively based on this, which can not only accurately focus on each feature, but also refer to different scale features at the same time, avoid information loss as much as possible, and comprehensively capture the deep structure of the insurance data to be tested, which is helpful to improve the accuracy of insurance fraud detection. Therefore, the efficiency and accuracy of insurance fraud detection can be improved, and the cost of insurance fraud detection can be reduced.

[0044] In some disclosed embodiments, the weight prediction module 32 includes a pooling submodule for performing global average pooling based on the first fusion feature to obtain a pooled reduced dimension feature; the weight prediction module 32 includes a prediction submodule for performing weight prediction based on the first fusion feature to obtain a local weight representation, and performing weight prediction based on the pooled reduced dimension feature to obtain a global weight representation; wherein the weight prediction includes at least one-dimensional convolution and batch normalization performed sequentially; the weight prediction module 32 includes a fusion submodule for performing fusion based on the global weight representation and the local weight representation to obtain a first weight representation; wherein the second weight representation is negatively correlated with the first weight representation.

[0045] In some disclosed embodiments, weight prediction includes sequentially performing one-dimensional convolution, batch normalization, linear rectification activation, one-dimensional convolution, and batch normalization.

[0046] In some disclosed embodiments, the fusion submodule includes an addition unit, which is used to perform element-by-element addition based on the global weight representation and the local weight representation to obtain a fused weight representation; perform element-by-element addition based on the global weight representation and the local weight representation to obtain a fused weight representation; the fusion submodule includes a constraint unit, which is used to constrain each element in the fused weight representation to a preset weight range based on a preset activation function to obtain a first weight representation.

[0047] In some disclosed embodiments, the insurance fraud detection device 30 includes a number detection module for detecting whether the first number of weighted fusions of the first scale feature and the second scale feature reaches a first threshold; the insurance fraud detection device 30 includes a feature selection module for selecting the second fused feature as a new first fused feature in response to the first number not reaching the first threshold, and returning to the step of predicting the first weight representation and the second weight representation based on the first fused feature until the first number reaches the first threshold.

[0048] In some disclosed embodiments, the feature extraction module 31 includes an initial extraction submodule for extracting initial data features of the insurance data to be tested; the feature extraction module 31 includes a first extraction submodule for performing a first extraction based on the initial data features to obtain a first scale feature; the feature extraction module 31 includes a second extraction submodule for performing a second extraction based on the initial data features to obtain a second scale feature; wherein the convolution kernel size used in the first extraction is smaller than the convolution kernel size used in the second extraction; the feature extraction module 31 includes a multi-scale fusion submodule for fusing the first scale feature and the second scale feature to obtain a multi-scale feature; the feature extraction module 31 includes a convolution processing submodule for performing at least convolution processing based on the multi-scale feature to obtain a first fusion feature.

[0049] In some disclosed embodiments, the first extraction includes sequentially performing one-dimensional convolution, batch normalization, pooling, and linear rectification activation; and / or, the second extraction includes sequentially performing one-dimensional convolution, batch normalization, pooling, linear rectification activation, one-dimensional convolution, batch normalization, pooling, and linear rectification activation; and / or, the multi-scale feature sequentially performs at least one-dimensional convolution, batch normalization, pooling, linear rectification activation, one-dimensional convolution, linear rectification activation, and full connection to obtain a first fused feature.

[0050] In some disclosed embodiments, the initial extraction submodule includes an initial encoding unit, which is used to perform one-hot encoding on discrete values ​​in the insurance data to be tested to obtain a first encoding feature, and to normalize and standardize continuous values ​​in the insurance data to be tested to obtain a second encoding feature; the initial extraction submodule includes a feature extraction unit, which is used to perform feature extraction based on the first encoding feature to obtain a first initial feature, and to perform feature extraction based on the second encoding feature to obtain a second initial feature; the initial extraction submodule includes a feature fusion unit, which is used to fuse the first initial feature and the second initial feature to obtain the initial data feature of the insurance data to be tested.

[0051] In some disclosed embodiments, the feature extraction module 31 includes a times detection submodule, which is used to detect whether the second times of performing multi-scale fusion of the first scale feature and the second scale feature reaches a second threshold; the feature extraction module 31 includes a multiple fusion submodule, which is used to select the multi-scale feature as the new initial data feature in response to the second times not reaching the second threshold, and return to the step of performing the first extraction based on the initial data feature to obtain the first scale feature, until the second times reaches the second threshold, select the latest first scale feature as the final first scale feature, and select the latest second scale feature as the final second scale feature.

[0052] In some disclosed embodiments, at least one of the first scale feature, the second scale feature, the first fusion feature, and the second fusion feature is obtained by a neural network model, and classification prediction is performed by a classifier based on ensemble learning.

[0053] See also Figure 4 , Figure 4 : is a schematic diagram of a framework of an embodiment of an electronic device of the present application. The electronic device 40 includes at least a memory 41 and a processor 42 coupled to each other, the memory 41 stores at least program instructions, and the processor 42 is used to execute the program instructions to implement the steps in any of the above-mentioned insurance fraud detection method embodiments. For details, please refer to the aforementioned disclosed embodiments, which will not be repeated here. As a possible example, the electronic device 40 may include but is not limited to a server, etc., and the specific type of the electronic device 40 is not limited here.

[0054] Specifically, the processor 42 is used to control itself and the memory 41 to implement the steps in any of the above-mentioned insurance fraud detection method embodiments. The processor 42 can also be called a CPU (Central Processing Unit). The processor 42 may be an integrated circuit chip with signal processing capabilities. The processor 42 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 42 can be implemented by an integrated circuit chip.

[0055] In the above scheme, the electronic device 40 extracts features based on the insurance data to be tested, obtains the first scale feature, the second scale feature and the first fusion feature of the first scale feature and the second scale feature, and predicts the first weight representation and the second weight representation based on the first fusion feature, and then obtains the second fusion feature based on the weighted feature of the first weight representation on the first scale feature and the weighted feature of the second weight representation on the second scale feature, and then performs classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested, and the detection result at least includes whether the insurance data to be tested is suspected of fraud. Therefore, on the one hand, since there is no need to manually identify the insurance data to be tested, the cost of insurance fraud detection can be reduced and the efficiency of insurance fraud detection can be improved. On the other hand, by performing weight prediction on the first fusion feature that integrates multiple scale features, the prediction accuracy of the reference degree of different scale features can be improved. Then, different scales are weighted respectively based on this, which can not only accurately focus on each feature, but also refer to different scale features at the same time, avoid information loss as much as possible, and comprehensively capture the deep structure of the insurance data to be tested, which helps to improve the accuracy of insurance fraud detection. Therefore, the efficiency and accuracy of insurance fraud detection can be improved, and the cost of insurance fraud detection can be reduced.

[0056] See also Figure 5 , Figure 5 The schematic diagram is a schematic diagram of a computer-readable storage medium 50 of the present application. The computer-readable storage medium 50 stores program instructions 51 that can be executed by a processor, and the program instructions 51 are used to implement the steps in any of the above-mentioned insurance fraud detection method embodiments.

[0057] In the above scheme, the computer-readable storage medium 50 extracts features based on the insurance data to be tested, obtains the first scale feature, the second scale feature, and the first fusion feature of the first scale feature and the second scale feature, and predicts the first weight representation and the second weight representation based on the first fusion feature, and then obtains the second fusion feature based on the weighted feature of the first weight representation on the first scale feature and the weighted feature of the second weight representation on the second scale feature, and then performs classification prediction based on the second fusion feature to obtain the detection result of the insurance data to be tested, and the detection result at least includes whether the insurance data to be tested is suspected of fraud. Therefore, on the one hand, since there is no need to manually identify the insurance data to be tested, the cost of insurance fraud detection can be reduced and the efficiency of insurance fraud detection can be improved. On the other hand, by performing weight prediction on the first fusion feature that integrates multiple scale features, the prediction accuracy of the reference degree of different scale features can be improved. Then, different scales are weighted respectively, which can not only accurately focus on each feature, but also refer to different scale features at the same time, avoid information loss as much as possible, and comprehensively capture the deep structure of the insurance data to be tested, which helps to improve the accuracy of insurance fraud detection. Therefore, the efficiency and accuracy of insurance fraud detection can be improved, and the cost of insurance fraud detection can be reduced.

[0058] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0059] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0060] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0061] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0062] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0063] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0064] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A method for detecting insurance fraud, characterized in that: include: Perform feature extraction based on the insurance data to be tested to obtain a first scale feature, a second scale feature, and a first fusion feature of the first scale feature and the second scale feature; Based on the first fusion feature, predict a first weighted representation and a second weighted representation; Obtain a second fused feature based on a weighted feature of the first scale feature represented by the first weight and a weighted feature of the second scale feature represented by the second weight; Classification prediction is performed based on the second fusion feature to obtain a detection result of the insurance data to be tested; wherein the detection result at least includes whether the insurance data to be tested is suspected of fraud.

2. The method according to claim 1, characterized in that The predicting, based on the first fusion feature, to obtain a first weighted representation and a second weighted representation includes: Perform global average pooling based on the first fusion feature to obtain a pooled dimensionality reduction feature; Performing weight prediction based on the first fused feature to obtain a local weight representation, and performing weight prediction based on the pooled dimensionality reduction feature to obtain a global weight representation; wherein the weight prediction includes at least one-dimensional convolution and batch normalization performed sequentially; The first weight representation is obtained by fusing the global weight representation and the local weight representation; wherein the second weight representation is negatively correlated with the first weight representation.

3. The method according to claim 2, characterized in that The weight prediction includes sequentially performing one-dimensional convolution, batch normalization, linear rectification activation, one-dimensional convolution, and batch normalization.

4. The method according to claim 2, characterized in that: The fusing the global weight representation and the local weight representation to obtain the first weight representation includes: Performing element-by-element addition based on the global weight representation and the local weight representation to obtain a fused weight representation; Based on a preset activation function, each element in the fusion weight representation is constrained to a preset weight range to obtain the first weight representation.

5. The method according to claim 1, characterized in that: After obtaining a second fused feature based on the first weighted feature of the first scale feature and the second weighted feature of the second scale feature by the first weight, and before performing classification prediction based on the second fused feature to obtain a detection result of the insurance data to be tested, the method further includes: detecting whether a first number of times the first scale feature and the second scale feature are weightedly fused reaches a first threshold; In response to the first number not reaching the first threshold, selecting the second fused feature as a new first fused feature, and returning to the step of predicting a first weighted representation and a second weighted representation based on the first fused feature until the first number reaches the first threshold.

6. The method according to claim 1, characterized in that The extracting features based on the insurance data to be tested to obtain the first scale feature, the second scale feature, and the first fusion feature of the first scale feature and the second scale feature includes: Extracting initial data features of the insurance data to be tested; Performing a first extraction based on the initial data feature to obtain the first scale feature, and performing a second extraction based on the initial data feature to obtain the second scale feature; wherein the convolution kernel size used in the first extraction is smaller than the convolution kernel size used in the second extraction; Based on the first scale feature and the second scale feature, fusing to obtain a multi-scale feature; At least convolution processing is performed based on the multi-scale features to obtain the first fusion features.

7. The method according to claim 6, characterized in that The first extraction includes sequentially performing one-dimensional convolution, batch normalization, pooling, and linear rectification activation; And / or, the second extraction includes sequentially performing one-dimensional convolution, batch normalization, pooling, linear rectification activation, one-dimensional convolution, batch normalization, pooling, linear rectification activation; And / or, the multi-scale features at least sequentially perform one-dimensional convolution, batch normalization, pooling, linear rectification activation, one-dimensional convolution, linear rectification activation and full connection to obtain the first fusion feature.

8. The method according to claim 6, characterized in that The extracting of initial data features of the insurance data to be tested includes: Performing one-hot encoding on discrete values ​​in the insurance data to be tested to obtain a first encoding feature, and normalizing and standardizing continuous values ​​in the insurance data to be tested to obtain a second encoding feature; Performing feature extraction based on the first coding feature to obtain a first initial feature, and performing feature extraction based on the second coding feature to obtain a second initial feature; The initial data features of the insurance data to be tested are obtained by fusing the first initial features and the second initial features.

9. The method according to claim 6, characterized in that After fusing the first scale feature and the second scale feature to obtain a multi-scale feature, and before performing at least convolution processing on the multi-scale feature to obtain the first fused feature, the method further includes: detecting whether a second number of times the first scale feature and the second scale feature perform multi-scale fusion reaches a second threshold; In response to the second number not reaching the second threshold, selecting the multi-scale feature as a new initial data feature, and returning to the step of performing the first extraction based on the initial data feature to obtain the first scale feature, until the second number reaches the second threshold, selecting the latest first scale feature as the final first scale feature, and selecting the latest second scale feature as the final second scale feature.

10. The method according to any one of claims 1 to 9, characterized in that: At least one of the first scale feature, the second scale feature, the first fusion feature, and the second fusion feature is obtained by a neural network model, and the classification prediction is performed by a classifier based on ensemble learning.

11. An insurance fraud detection device, characterized in that: include: A feature extraction module, used for performing feature extraction based on the insurance data to be tested, and obtaining a first scale feature, a second scale feature, and a first fusion feature of the first scale feature and the second scale feature; A weight prediction module, used for performing prediction based on the first fusion feature to obtain a first weight and a second weight; A weighted fusion module, configured to obtain a second fused feature based on a first weighted feature of the first scale feature obtained by the first weight and a second weighted feature of the second scale feature obtained by the second weight; A classification prediction module is used to perform classification prediction based on the second fusion feature to obtain a detection result of the insurance data to be tested; wherein the detection result at least includes whether the insurance data to be tested is suspected of fraud.

12. An electronic device, characterized in that: The invention comprises at least a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the insurance fraud detection method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: Program instructions that can be executed by a processor are stored, and the program instructions are used to implement the insurance fraud detection method according to any one of claims 1 to 10.

Images