Deep learning-based insurance abnormal behavior identification method and system

Abstract

The invention relates to the field of intelligent data analysis, and provides an insurance abnormal behavior identification method and system based on deep learning, and the method employs a data processing technology based on deep learning to carry out the embedded coding of a data set of personal communication information, and then carries out the full-time-domain cascading. And performing significant aggregation between embedded coding and feature effects on the data set of the doctor-seeing position information, so as to guide significant interaction representation according to conditions between communication embedded cascade features and doctor-seeing position significant aggregation representation to intelligently judge whether the doctor-seeing behavior of the target user is abnormal or not. In this way, complex interaction among different features can be captured, and therefore abnormal behaviors can be recognized more accurately. And meanwhile, the system can be continuously optimized along with the evolution of fraud means, so that better dynamic adaptability is achieved to cope with continuously changing fraud behaviors, and the accuracy and the intelligent degree of insurance abnormal behavior auditing are improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent data analysis, and more specifically, to a method and system for identifying abnormal insurance behavior based on deep learning. Background Art

[0002] The insurance industry has long faced the threat of fraud, especially in the field of medical insurance. Fraud not only causes financial losses to insurance companies, but may also lead to rising insurance premiums, affecting the healthy development of the entire industry. Therefore, the development of effective fraud detection technology has become an urgent need for the insurance industry.

[0003] Patent CN118799089A proposes a method, device, equipment and storage medium for identifying insurance information anomalies. Specifically, after receiving an insurance review request, the system determines the user's identity and analyzes their communication information, evaluates the authenticity and consistency of the user's medical behavior and location, and gives a score based on this evaluation for the insurance agency to determine whether the information is abnormal.

[0004] In the above patent, the identification of abnormal insurance behavior is completed by analyzing the user's personal communication information and the medical details in the insurance review information, calculating the medical location verification score and the medical behavior verification score, and combining these two scores to obtain the medical behavior evaluation score. However, this rule-based method in the patent is rigid in dealing with evolving fraud methods because it relies on predefined rules and fixed weights, and it is difficult to adjust in time to capture new fraud patterns. In addition, this type of method is limited by linear logic and often ignores nonlinear interactions between features, resulting in an inability to fully assess risks in complex situations. For example, in a dynamically changing environment, new types of fraud may appear and spread rapidly, and rule-driven systems require manual rule updates to respond, which is often lagging and inflexible in actual operations. At the same time, failure to fully consider the interactions between features means that important risk signals may be missed, reducing the accuracy and reliability of detection.

[0005] Therefore, an optimized insurance abnormal behavior identification solution is desired. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present application provides a method and system for identifying abnormal insurance behavior based on deep learning.

[0007] According to one aspect of the present application, a method for identifying abnormal insurance behavior based on deep learning is provided, which includes: receiving insurance audit information sent by an insurance institution; extracting a data set of personal communication information and a data set of medical location information of a target user from the insurance audit information; based on the data set of personal communication information and the data set of medical location information, performing abnormal identification on the medical behavior of the target user to obtain an abnormal identification result of the insurance audit information, including: embedding and coding the data set of personal communication information and the data set of medical location information to obtain a time series of communication embedded coding features and a time series of medical location embedded coding features; performing significant aggregation of inter-feature forces on the time series of medical location embedded coding features to obtain a significant aggregation representation of the medical location; cascading the time series of communication embedded coding features to obtain communication embedded cascade features, and performing conditional information-guided significant interaction with the significant aggregation representation of the medical location to obtain a significant interaction fusion representation of the medical location-communication features; based on the significant interaction fusion representation of the medical location-communication features, obtaining the abnormal identification result, and the abnormal identification result is used to indicate whether the medical behavior of the target user is abnormal.

[0008] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the dataset of the personal communication information and the dataset of the medical visit location information are embedded and encoded to obtain a time series of communication embedded coding features and a time series of medical visit location embedded coding features, including: embedding and encoding the dataset of the personal communication information through a communication embedding matrix to obtain a time series of communication embedded coding vectors as the time series of the communication embedded coding features; embedding and encoding the dataset of the medical visit location information through a medical visit location embedding matrix to obtain a time series of medical visit location embedded coding vectors as the time series of the medical visit location embedded coding features.

[0009] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the time series of the embedded coding features of the visit location is subjected to significant aggregation of the forces between the features to obtain a significantly aggregated representation of the visit location, including: calculating the static energy factors of each visit location embedded coding vector in the time series of the visit location embedded coding vector to obtain a sequence of visit location static energy factors; selecting the visit location embedded coding vector corresponding to the largest visit location static energy factor in the sequence of visit location static energy factors as the initial center vector of the visit location clustering; based on the sequence of visit location static energy factors and the initial center vector of the visit location clustering, performing energy distribution-based feature aggregation on the time series of the visit location embedded coding vector to obtain a visit location significantly aggregated representation vector as the visit location significantly aggregated representation.

[0010] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the static energy factors of each visit location embedded coding vector in the time series of the visit location embedded coding vector are calculated to obtain a sequence of visit location static energy factors, including: calculating the mean and variance of the visit location embedded coding vector to obtain the visit location mean and the visit location variance; subtracting the visit location embedded coding vector from the visit location mean by position, and calculating the fourth power of each position of the subtracted feature vector to obtain the visit location difference modulation vector; calculating the expected value of the visit location difference modulation vector to obtain the visit location expected value; dividing the visit location expected value by the square of the visit location variance, and inputting the divided value into the sigmoid function to obtain the visit location static energy factor.

[0011] In the above-mentioned insurance abnormal behavior identification method based on deep learning, based on the sequence of static energy factors of the visit location and the initial center vector of the visit location cluster, the time series of the visit location embedded coding vector is subjected to feature aggregation based on energy distribution to obtain a visit location significantly aggregated representation vector, including: based on the spatial span between each visit location embedded coding vector in the time series of the visit location embedded coding vector and the initial center vector of the visit location cluster and the static energy factor of each visit location embedded coding vector and the static energy factor of the initial center vector of the visit location cluster, the dynamic aggregation energy factor of each visit location embedded coding vector in the time series of the visit location embedded coding vector is calculated to obtain a sequence of visit location dynamic aggregation energy factors; the sequence of visit location dynamic aggregation energy factors is input into a gated mask unit to obtain a sequence of visit location dynamic aggregation weight factors; and the weighted sum of the time series of the visit location embedded coding vector is calculated using the sequence of visit location dynamic aggregation weight factors to obtain the visit location significantly aggregated representation vector.

[0012] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the time series of the communication embedded coding features is cascaded to obtain the communication embedded cascade features, and then the conditional information is guided to significantly interact with the medical location significant aggregation representation to obtain the medical location-communication feature significant interaction fusion representation, including: cascading the time series of the communication embedded coding vectors to obtain the communication embedded cascade vector as the communication embedded cascade feature; conditional information is guided to significantly interact with the medical location significant aggregation representation vector and the communication embedded cascade vector to obtain the medical location-communication feature significant interaction fusion representation vector as the medical location-communication feature significant interaction fusion representation.

[0013] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the significantly aggregated representation vector of the medical location and the communication embedded cascade vector are significantly interacted under the guidance of conditional information to obtain a significantly interactive fusion representation vector of the medical location-communication feature, including: inputting the significantly aggregated representation vector of the medical location and the communication embedded cascade vector into a joint implicit feature capture network to obtain a joint implicit feature vector of the medical location-communication context; performing feature activation based on the Sigmoid function on the joint implicit feature vector of the medical location-communication context to obtain a medical location-communication condition feature vector; calculating the medical location semantic contribution of the significantly aggregated representation vector of the medical location relative to the medical location-communication condition feature vector; calculating the communication semantic contribution of the communication embedded cascade vector to the medical location-communication condition feature vector. ; The semantic contribution of the medical location and the semantic contribution of the communication are normalized, and the normalized semantic contribution of the medical location and the normalized communication semantic contribution are used to perform weighted modulation on the medical location significant aggregation representation vector and the communication embedding cascade vector to obtain the modulated medical location significant aggregation representation vector and the modulated communication embedding cascade vector; using the modulated medical location significant aggregation representation vector as the query vector, the modulated communication embedding cascade vector as the key vector and the medical location-communication condition feature vector as the value vector, the modulated medical location significant aggregation representation vector, the modulated communication embedding cascade vector and the medical location-communication condition feature vector are input into the feature-significant guided interaction module based on the converter structure to obtain the medical location-communication feature significant interaction fusion representation vector.

[0014] In the above-mentioned insurance abnormal behavior identification method based on deep learning, the significant aggregation representation vector of the medical location and the communication embedding cascade vector are input into the joint implicit feature capture network to obtain the medical location-communication context joint implicit feature vector, including: adding the significant aggregation representation vector of the medical location and the communication embedding cascade vector by position, multiplying the obtained medical location-communication sum vector with the weight matrix and then adding it with the bias vector by position to obtain the medical location-communication joint interaction vector; using the tanh function to process the medical location-communication joint interaction vector to obtain the medical location-communication context joint implicit feature vector.

[0015] In the above-mentioned deep learning-based insurance abnormal behavior identification method, the abnormal identification result is obtained based on the significant interaction fusion representation of the medical location-communication features, including: passing the medical location-communication features significant interaction fusion representation vector through a classifier-based medical behavior abnormality identifier to obtain the abnormal identification result, and the abnormal identification result is used to indicate whether the medical behavior of the target user is abnormal.

[0016] According to another aspect of the present application, a deep learning-based insurance abnormal behavior identification system is provided, which includes: an insurance review information receiving module, used to receive insurance review information sent by an insurance institution; a personal communication and medical location information extraction module, used to extract a data set of personal communication information and a data set of medical location information of a target user from the insurance review information; an insurance review information anomaly identification module, used to identify the anomaly of the medical behavior of the target user based on the data set of personal communication information and the data set of medical location information to obtain an anomaly identification result of the insurance review information, wherein the insurance review information anomaly identification module includes: a personal communication and medical location information embedding encoding unit, used to embed the data set of personal communication information and the data set of medical location information Encoding to obtain a time series of communication embedded coding features and a time series of visit location embedded coding features; a visit location significant aggregation representation generation unit, used to perform significant aggregation of the inter-feature forces on the time series of visit location embedded coding features to obtain a visit location significant aggregation representation; a visit location-communication feature significant interaction fusion representation generation unit, used to cascade the time series of communication embedded coding features to obtain communication embedded cascade features and then perform conditional information guided significant interaction with the visit location significant aggregation representation to obtain a visit location-communication feature significant interaction fusion representation; a visit behavior anomaly identification unit, used to obtain the anomaly identification result based on the visit location-communication feature significant interaction fusion representation, and the anomaly identification result is used to indicate whether the target user's visit behavior is abnormal.

[0017] The present application has significant technical effects due to the adoption of the above technical solutions: the deep learning-based insurance abnormal behavior identification method and system provided by the present application adopts a data processing technology based on deep learning to embed the data set of personal communication information and then cascade it in the entire time domain, and embed the data set of medical location information and perform significant aggregation between feature effects, so as to intelligently judge whether the medical behavior of the target user is abnormal based on the conditional guidance of significant interaction representation between the communication embedded cascade feature and the significant aggregation representation of the medical location. In this way, the complex interactions between different features can be captured, so as to more accurately identify abnormal behaviors. At the same time, it can continuously optimize itself as the means of fraud evolve, so as to have better dynamic adaptability to cope with the ever-changing fraudulent behaviors, and improve the accuracy and intelligence of the review of abnormal insurance behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 This is a flowchart of a method for identifying abnormal insurance behavior based on deep learning according to an embodiment of the present application.

[0020] Figure 2 In the deep learning-based insurance abnormal behavior identification method according to an embodiment of the present application, based on the data set of the personal communication information and the data set of the medical location information, the flowchart of performing abnormal identification on the medical behavior of the target user to obtain the abnormal identification result of the insurance review information.

[0021] Figure 3 In the deep learning-based insurance abnormal behavior identification method according to an embodiment of the present application, based on the data set of the personal communication information and the data set of the medical treatment location information, the data flow diagram of the abnormal identification of the target user's medical treatment behavior is performed to obtain the abnormal identification result of the insurance review information.

[0022] Figure 4 The present invention is a flowchart of performing significant aggregation of inter-feature forces on the time series of embedded coding features of the medical consultation locations in the insurance abnormal behavior identification method based on deep learning according to an embodiment of the present application to obtain a significant aggregation representation of the medical consultation locations.

[0023] Figure 5 In the deep learning-based insurance abnormal behavior identification method according to an embodiment of the present application, the time series of the communication embedded coding features is cascaded to obtain the communication embedded cascade features, and then the conditional information is used to guide significant interaction with the significant aggregation representation of the medical location to obtain a flowchart of the significant interaction fusion representation of the medical location-communication features.

[0024] Figure 6 This is a system block diagram of an insurance abnormal behavior identification system based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0026] It should be noted that in this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization of the owner of the corresponding device.

[0027] The insurance industry has long faced the challenge of fraud, especially in the field of medical insurance. Patent CN118799089A proposes a method, device, equipment and storage medium for identifying anomalies in insurance information. Specifically, by analyzing the user's communication information, the authenticity and consistency of the user's medical behavior and location are evaluated and scored accordingly, so that insurance institutions can determine whether the information is abnormal. However, this method relies on predefined rules and fixed weights, and it is difficult to adapt to the evolution of fraud methods. It is limited by linear logic and ignores nonlinear interactions between features. It cannot fully assess risks in complex situations. For example, when new types of fraud appear, the rules need to be manually updated, which has lags and may miss important risk signals, reducing detection accuracy and reliability.

[0028] Based on this, this application proposes a method for identifying abnormal insurance behavior based on deep learning. Figure 1 FIG. 1 is a flow chart of a method for identifying abnormal insurance behavior based on deep learning according to an embodiment of the present application. Figure 1 As shown, according to the deep learning-based insurance abnormal behavior identification method of the embodiment of the present application, the method includes: S110, receiving insurance audit information sent by an insurance institution; S120, extracting a data set of personal communication information and a data set of medical treatment location information of a target user from the insurance audit information; S130, based on the data set of personal communication information and the data set of medical treatment location information, performing abnormal identification on the medical treatment behavior of the target user to obtain an abnormal identification result of the insurance audit information.

[0029] In step S110, the insurance review information sent by the insurance agency is received. The insurance agency is an agency for users to apply for insurance claims. The insurance review information includes but is not limited to the key application information of the user who applies for insurance claims and the longitude and latitude coordinate information of the hospital where the patient is treated. The key application information is integrated by the insurance agency from the various application materials submitted by the user, including but not limited to the user's contact information, name, consultation time, length of hospitalization, name of the hospital where the patient is treated, type of medical treatment, condition, time of medical treatment and other basic information. After receiving the various application materials submitted by the user, the insurance agency can call the public map interface data to obtain the longitude and latitude coordinate information of the hospital where the user is treated at different consultation times. The user's insurance review information is finally formed by sorting out the user's key application information and the longitude and latitude coordinate information of the hospital where the patient is treated.

[0030] In step S120, a data set of personal communication information and a data set of medical treatment location information of the target user are extracted from the insurance review information. It should be understood that the user's personal communication information, such as the location signaling record of the mobile phone signal, call data, etc., can understand the area where the user usually moves, including their permanent location during the day and at night. The medical treatment location information includes the geographical location of the hospital where the user is when seeking medical treatment. By analyzing the user's communication records and medical treatment location, it can be verified whether the user actually seeks medical treatment in the declared hospital. If the user's communication records show that they did not appear near the hospital during the time of medical treatment, then this may indicate the existence of false claims. In addition, if the user's behavior pattern (such as frequent changes of location) does not match their medical records, or they seek medical treatment multiple times in a short period of time, this may be a sign of fraud. Therefore, combining personal communication information with medical treatment location information can effectively assist insurance institutions in making risk judgments when processing medical insurance claims, ensuring that only real medical expenses are compensated, thereby maintaining the healthy operation of the insurance system. In particular, in an embodiment of the present application, a possible implementation method for extracting a data set of personal communication information and a data set of medical location information of the target user from the insurance review information may be: according to the key application information in the insurance review information, determine that the user who applies for insurance claims is the target user, and retrieve the target user's personal communication information including call data and mobile phone signal location signaling data at different medical visits from the operator according to the contact information of the target user in the insurance review information, and organize these data in chronological order to obtain a data set of personal communication information. The insurance review data already contains the latitude and longitude coordinate information of different medical hospitals for users sorted by the insurance agency. The medical location-related information can be extracted from the insurance review information through regular expressions, and the medical location information data set can be obtained by sorting them according to time.

[0031] In step S130, based on the data set of the personal communication information and the data set of the medical treatment location information, abnormal identification is performed on the medical treatment behavior of the target user to obtain an abnormal identification result of the insurance review information.

[0032] Accordingly, in the process of performing abnormal identification on the medical treatment behavior of the target user based on the data set of the personal communication information and the data set of the medical treatment location information to obtain the abnormal identification result of the insurance audit information, the technical concept of the present application is to use data analysis and processing technology based on deep learning to embed the data set of the personal communication information and then cascade it in the entire time domain, and embed the data set of the medical treatment location information and perform significant aggregation between feature effects, so as to intelligently judge whether the medical treatment behavior of the target user is abnormal based on the conditional guidance of significant interaction representation between the communication embedded cascade feature and the significant aggregation representation of the medical treatment location. In this way, the complex interactions between different features can be captured, rather than just simple linear relationships, so as to more accurately identify abnormal behaviors. At the same time, it can continuously optimize itself as the means of fraud evolve, so as to have better dynamic adaptability to cope with the ever-changing fraudulent behaviors, thereby improving the accuracy and intelligence of the review of abnormal insurance behaviors.

[0033] Specifically, Figure 2 In the deep learning-based insurance abnormal behavior identification method according to an embodiment of the present application, based on the data set of the personal communication information and the data set of the medical location information, the abnormal identification of the medical behavior of the target user is performed to obtain the abnormal identification result of the insurance review information. Figure 3 This is a data flow diagram of the abnormal insurance behavior identification method based on deep learning according to an embodiment of the present application, in which the abnormal medical behavior of the target user is identified based on the data set of the personal communication information and the data set of the medical location information to obtain the abnormal identification result of the insurance review information. Figure 2 and Figure 3 As shown, based on the data set of the personal communication information and the data set of the medical location information, the medical treatment behavior of the target user is identified as abnormal to obtain the abnormal identification result of the insurance review information, including: S131, embedding and coding the data set of the personal communication information and the data set of the medical location information to obtain the time series of communication embedded coding features and the time series of medical location embedded coding features; S132, performing significant aggregation of the inter-feature forces on the time series of the medical location embedded coding features to obtain a significant aggregation representation of the medical location; S133, cascading the time series of the communication embedded coding features to obtain the communication embedded cascade feature, and then performing conditional information-guided significant interaction with the medical location significant aggregation representation to obtain a medical location-communication feature significant interaction fusion representation; S134, based on the medical location-communication feature significant interaction fusion representation, obtaining the abnormal identification result, and the abnormal identification result is used to indicate whether the medical treatment behavior of the target user is abnormal.

[0034] In step S131, embedding coding is performed on the dataset of the personal communication information and the dataset of the medical consultation location information to obtain a time series of communication embedded coding features and a time series of medical consultation location embedded coding features. Specifically, in an embodiment of the present application, embedding coding is performed on the dataset of the personal communication information and the dataset of the medical consultation location information to obtain a time series of communication embedded coding features and a time series of medical consultation location embedded coding features, including: embedding coding the dataset of the personal communication information through a communication embedding matrix to obtain a time series of communication embedded coding vectors as a time series of communication embedded coding features; embedding coding the dataset of the medical consultation location information through a medical consultation location embedding matrix to obtain a time series of medical consultation location embedded coding vectors as a time series of medical consultation location embedded coding features.

[0035] It should be understood that the data set of the personal communication information may be high-dimensional and sparse and have complex spatiotemporal correlations, such as discrete values ​​such as mobile phone numbers and base station IDs. Directly using these original values ​​will increase the complexity of the model and make it difficult to train. Based on this, in order to better understand and predict the behavior patterns and internal connections of users at different time points, in the technical solution of the present application, the data set of the personal communication information is embedded and encoded through a communication embedding matrix to obtain a time series of communication embedding coding vectors. In this way, the personal communication information containing time dimension information can be converted into a sequence of vectors, and each vector represents the state of the communication information at a time point, so that the dynamic change trend of the communication information over time can be better characterized.

[0036] Similarly, considering that the dataset of the medical location information contains the location data of the hospitals where the target user visited at different times, the high-dimensional sparse expression formed by a large amount of different location data will make the subsequent calculations highly complex and may cause the model to overfit. In order to process high-dimensional sparse data, in the technical solution of the present application, the dataset of the medical location information is embedded and encoded through a medical location embedding matrix to obtain a time series of medical location embedded coding vectors. By embedding and encoding the dataset of the medical location information, the original high-dimensional sparse data can be mapped to a low-dimensional dense space, thereby making the medical location data easier to process. Moreover, each vector in the time series of medical location embedded coding vectors obtained by embedding coding represents the medical location information state at a medical time. This sequence can well characterize the dynamic changes of the medical location with the medical time.

[0037] In step S132, the time series of the embedded coding features of the clinic location is subjected to significant aggregation of the forces between the features to obtain a significantly aggregated representation of the clinic location. Accordingly, considering that different embedded coding features of the clinic location in the time series of the embedded coding vector of the clinic location represent a feature of the clinic location information, and different clinic location information has different emphases and importance in the subsequent identification of abnormal clinic behavior, and each clinic location has a different influence on the characteristics of the entire central behavior. Therefore, in order to enhance the importance representation of these key features and to highlight those features that have a greater influence on anomaly detection, in the technical solution of the present application, the time series of the embedded coding vector of the clinic location is subjected to significant aggregation of the forces between the features to obtain a significantly aggregated representation vector of the clinic location as the significantly aggregated representation of the clinic location.

[0038] Specifically, first, the static energy factor of each clinic location embedded coding vector in the time series of the clinic location embedded coding vector is calculated to evaluate the relative importance of each feature vector in the data set, obtain a sequence of clinic location static energy factors, and provide a reference for determining the initial clustering center. Next, in order to select the most representative features from the entire data set to enhance the ability to capture the features of the clustering process, the clinic location embedded coding vector corresponding to the largest clinic location static energy factor is selected as the initial center vector of the clinic location clustering. Then, based on the spatial span between each and the initial center vector, the dynamic aggregation energy factor of the clinic location is calculated for each clinic location static energy factor to more accurately reveal the complex relationship and intrinsic importance between samples through more sophisticated interactive calculations. Afterwards, in order to automatically adjust the weight of each energy factor according to the importance of the learned features, the dynamic aggregation energy factor of each clinic location is input into the gated mask unit to enhance the focus on key features, and a sequence of dynamic aggregation weight factors of the clinic location is obtained. Finally, based on the dynamic aggregation weight factors of each visit location in the sequence of dynamic aggregation weight factors of the visit locations, the time series of the visit location embedded coding vectors are weighted and calculated to selectively enhance or weaken them according to their importance, thereby forming a more concentrated and representative expression, namely the so-called visit location significant aggregation representation vector.

[0039] Specifically, Figure 4 The present invention is a flowchart of performing significant aggregation of the inter-feature forces on the time series of the embedded coding features of the medical consultation location in the deep learning-based insurance abnormal behavior identification method according to an embodiment of the present application to obtain a significant aggregation representation of the medical consultation location. Figure 4As shown, the time series of the embedded coding features of the visit locations is subjected to significant aggregation of the forces between the features to obtain a significantly aggregated representation of the visit locations, including: S1321, calculating the static energy factors of each visit location embedded coding vector in the time series of the visit location embedded coding vector to obtain a sequence of the visit location static energy factors; S1322, selecting the visit location embedded coding vector corresponding to the largest visit location static energy factor in the sequence of the visit location static energy factors as the initial center vector of the visit location clustering; S1323, based on the sequence of the visit location static energy factors and the initial center vector of the visit location clustering, performing energy distribution-based feature aggregation on the time series of the visit location embedded coding vector to obtain a visit location significantly aggregated representation vector as the visit location significantly aggregated representation.

[0040] More specifically, in an embodiment of the present application, the static energy factors of each visit position embedded coding vector in the time series of the visit position embedded coding vector are calculated to obtain a sequence of visit position static energy factors, including: calculating the mean and variance of the visit position embedded coding vector to obtain the visit position mean and the visit position variance; subtracting the visit position embedded coding vector from the visit position mean by position, and calculating the fourth power of each position of the subtracted eigenvector to obtain the visit position difference modulation vector; calculating the expected value of the visit position difference modulation vector to obtain the visit position expected value; dividing the visit position expected value by the square of the visit position variance, and inputting the divided value into the sigmoid function to obtain the visit position static energy factor.

[0041] More specifically, in an embodiment of the present application, based on the sequence of static energy factors of the clinic location and the initial center vector of the clinic location cluster, the time series of the clinic location embedded coding vectors is subjected to feature aggregation based on energy distribution to obtain a clinic location significantly aggregated representation vector, including: based on the spatial span between each clinic location embedded coding vector in the time series of the clinic location embedded coding vector and the initial center vector of the clinic location cluster and the static energy factor of each clinic location embedded coding vector and the static energy factor of the initial center vector of the clinic location cluster, calculating the dynamic aggregation energy factor of each clinic location embedded coding vector in the time series of the clinic location embedded coding vector to obtain a sequence of clinic location dynamic aggregation energy factors; inputting the sequence of clinic location dynamic aggregation energy factors into a gated mask unit to obtain a sequence of clinic location dynamic aggregation weight factors; and calculating the weighted sum of the time series of the clinic location embedded coding vectors using the sequence of clinic location dynamic aggregation weight factors to obtain the clinic location significantly aggregated representation vector.

[0042] More specifically, in an embodiment of the present application, based on the spatial span between each visit location embedded coding vector in the time series of the visit location embedded coding vector and the initial center vector of the visit location cluster, as well as the static energy factor of each visit location embedded coding vector and the static energy factor of the initial center vector of the visit location cluster, the dynamic aggregation energy factor of each visit location embedded coding vector in the time series of the visit location embedded coding vector is calculated to obtain a sequence of visit location dynamic aggregation energy factors, including: multiplying the static energy factor of the visit location embedded coding vector and the static energy factor of the initial center vector of the visit location cluster with a first weighting parameter to obtain a first visit location dynamic aggregation energy factor; multiplying the square of the spatial span between the visit location embedded coding vector and the initial center vector of the visit location cluster with a second weighting parameter to obtain a second visit location dynamic aggregation energy factor; dividing the first visit location dynamic aggregation energy factor by the second visit location dynamic aggregation energy factor to obtain the visit location dynamic aggregation energy factor.

[0043] More specifically, in an embodiment of the present application, the sequence of dynamic aggregation energy factors of the visit locations is input into a gated mask unit to obtain a sequence of dynamic aggregation weight factors of the visit locations, including: using a sigmoid function to normalize the sequence of dynamic aggregation energy factors of the visit locations to obtain a sequence of normalized dynamic aggregation energy factors of the visit locations; and masking the sequence of normalized dynamic aggregation energy factors of the visit locations to obtain a sequence of dynamic aggregation weight factors of the visit locations.

[0044] In the embodiment of the present application, specifically, the time series of the embedded coding vector of the medical examination location is subjected to significant aggregation of the inter-feature forces, which can be expressed by the formula: ;in, is the time series of the embedding encoding vectors for the medical visit locations, are the first, second, ..., and third positions in the time series of the embedding coding vector of the medical consultation location. ,..., The patient location embedding encoding vector, It is The characteristic values ​​of each position in the embedding encoding vector of the medical location are To calculate the expected value, and They are The mean and variance of yes function, yes The corresponding static energy factor of the treatment location, To select the maximum value corresponding to value, is the maximum matching value, is the initial center vector of the clinic location cluster, yes The corresponding static energy factor of the treatment location, express and The space span between and is the weighting parameter, for The corresponding treatment location dynamic aggregation energy factor, yes The corresponding normalized visit location dynamic aggregation energy factor, For masking, is a predetermined threshold, yes The corresponding dynamic aggregation weight factor of the treatment location, is the number of vectors in the time series of the embedding coding vector of the medical consultation location, is the significant aggregation representation vector of the visit location.

[0045] In step S133, the time series of the communication embedding coding features are cascaded to obtain the communication embedding cascade features, and then the conditional information is guided to significantly interact with the clinic location significant aggregation representation to obtain the clinic location-communication feature significant interaction fusion representation. Specifically, Figure 5 In the deep learning-based insurance abnormal behavior identification method according to the embodiment of the present application, the time series of the communication embedding coding features are cascaded to obtain the communication embedding cascade features, and then the conditional information is guided to significantly interact with the significant aggregation representation of the medical treatment location to obtain the flow chart of the significant interaction fusion representation of the medical treatment location-communication features. Figure 5 As shown, after the time series of the communication embedding coding features are cascaded to obtain the communication embedding cascade features, the conditional information-guided significant interaction is performed with the significantly aggregated representation of the medical location to obtain the medical location-communication feature significant interaction fusion representation, including: S1331, cascading the time series of the communication embedding coding vectors to obtain the communication embedding cascade vector as the communication embedding cascade feature; S1332, the conditional information-guided significant interaction is performed on the significantly aggregated representation vector of the medical location and the communication embedding cascade vector to obtain the medical location-communication feature significant interaction fusion representation vector as the medical location-communication feature significant interaction fusion representation.

[0046] It should be understood that in order to integrate and retain the dynamic information in the time series of the communication embedded coding vector and form a complete feature representation, so as to better capture the time dependency of the user's communication behavior, in the technical solution of the present application, the time series of the communication embedded coding vector is cascaded to obtain a communication embedded cascade vector, thereby retaining the dynamic change characteristics in the time series of the communication information, that is, the trend of the user's communication behavior over time, so as to more comprehensively understand the user's communication characteristics.

[0047] Furthermore, considering that the significantly aggregated representation vector of the medical visit location and the communication embedded cascade vector respectively express the characteristics of different aspects of the user's behavior, in order to integrate and strengthen the interaction between data from different sources, so as to more accurately identify the user's abnormal behavior, in the technical solution of the present application, the significantly aggregated representation vector of the medical visit location and the communication embedded cascade vector are subjected to conditional information-guided significant interaction to obtain the medical visit location-communication feature significant interaction fusion representation vector.

[0048] In detail, first, the significant aggregation representation vector of the medical location and the communication embedding cascade vector are input into the joint implicit feature capture network to capture and learn the intrinsic correlation between the two sets of input features, and obtain the joint implicit feature vector of the medical location-communication context. Then, in order to further refine the implicit information of the two and give it appropriate weights, the Sigmoid function is used to perform feature activation processing on the joint implicit feature vector of the medical location-communication context to obtain the medical location-communication conditional feature vector. Then, the importance and influence of the significant aggregation representation vector of the medical location and the communication embedding cascade vector relative to the conditional feature vector are calculated respectively to obtain the semantic contribution of the medical location and the semantic contribution of the medical location. Subsequently, the two semantic contributions are normalized so that all their weights are on the same scale, and based on the normalized medical location semantic contribution and the normalized communication semantic contribution, the significant aggregation representation vector of the medical location and the communication embedding cascade vector are feature modulated to obtain the modulated significant aggregation representation vector of the medical location and the modulated communication embedding cascade vector. Finally, the modulated visit location significant aggregation representation vector is used as the query vector, the modulated communication embedding cascade vector is used as the key vector, and the visit location-communication condition feature vector is used as the value vector. They are input into the feature significant guided interaction module based on the converter structure to realize the interaction and fusion between different features. That is, the converter structure can better understand the contextual relationship between features, more effectively grasp the correlation between features, and automatically focus on the most critical features, and finally obtain the visit location-communication feature significant interaction fusion representation vector.

[0049] Specifically, in an embodiment of the present application, the significantly aggregated representation vector of the medical consultation location and the communication embedding cascade vector are subjected to conditional information-guided significant interaction to obtain a significantly interactive fusion representation vector of the medical consultation location-communication feature, including: inputting the significantly aggregated representation vector of the medical consultation location and the communication embedding cascade vector into a joint implicit feature capture network to obtain a joint implicit feature vector of the medical consultation location-communication context; performing feature activation based on the Sigmoid function on the joint implicit feature vector of the medical consultation location-communication context to obtain a medical consultation location-communication condition feature vector; calculating the medical consultation location semantic contribution of the significantly aggregated representation vector of the medical consultation location relative to the medical consultation location-communication condition feature vector; calculating the communication semantic contribution of the communication embedding cascade vector to the medical consultation location-communication condition feature vector; and The clinic location semantic contribution and the communication semantic contribution are normalized, and the normalized clinic location semantic contribution and the normalized communication semantic contribution are used to perform weighted modulation on the clinic location significant aggregation representation vector and the communication embedding cascade vector to obtain the modulated clinic location significant aggregation representation vector and the modulated communication embedding cascade vector; using the modulated clinic location significant aggregation representation vector as the query vector, the modulated communication embedding cascade vector as the key vector and the clinic location-communication condition feature vector as the value vector, the modulated clinic location significant aggregation representation vector, the modulated communication embedding cascade vector and the clinic location-communication condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain the clinic location-communication feature significant interaction fusion representation vector.

[0050] More specifically, in an embodiment of the present application, the visit location salient aggregation representation vector and the communication embedding cascade vector are input into a joint implicit feature capture network to obtain a visit location-communication context joint implicit feature vector, including: adding the visit location salient aggregation representation vector and the communication embedding cascade vector by position, multiplying the obtained visit location-communication sum vector with a weight matrix, and then adding the vector by position with a bias vector to obtain a visit location-communication joint interaction vector; and processing the visit location-communication joint interaction vector using a tanh function to obtain the visit location-communication context joint implicit feature vector.

[0051] More specifically, in an embodiment of the present application, the semantic contribution of the clinic location of the significant aggregation representation vector relative to the clinic location-communication condition feature vector is calculated, including: calculating the positional division of the corresponding positions of the clinic location significant aggregation representation vector and the clinic location-communication condition feature vector to obtain the clinic location semantic contribution vector; calculating the base-two logarithmic function value of the absolute value of each eigenvalue of the clinic location semantic contribution vector to obtain the clinic location semantic contribution logarithmic vector; calculating the positional dot product of the clinic location significant aggregation representation vector and the clinic location semantic contribution logarithmic vector, and performing position-by-position multiplication on the obtained dot product vector to obtain the clinic location semantic contribution value; calculating an exponential function with the natural constant e as the base and the clinic location semantic contribution value as the exponent to obtain the clinic location semantic contribution. Similarly, the processing method of the communication semantic contribution is consistent with the processing method of the clinic location semantic contribution.

[0052] More specifically, in an embodiment of the present application, the semantic contribution of the clinic location and the communication semantic contribution are normalized, and the normalized semantic contribution of the clinic location and the normalized communication semantic contribution are used to perform weighted modulation on the clinic location significant aggregation representation vector and the communication embedding cascade vector to obtain the modulated clinic location significant aggregation representation vector and the modulated communication embedding cascade vector, including: calculating the sum of the clinic location semantic contribution and the communication semantic contribution to obtain the clinic location-communication semantic contribution sum; dividing the clinic location semantic contribution and the communication semantic contribution by the clinic location-communication semantic contribution sum to obtain the normalized clinic location semantic contribution and the normalized communication semantic contribution; performing positional point multiplication of the clinic location significant aggregation representation vector and the normalized clinic location semantic contribution to obtain the modulated clinic location significant aggregation representation vector; performing positional point multiplication of the communication embedding cascade vector and the normalized communication semantic contribution to obtain the modulated communication embedding cascade vector.

[0053] In the embodiment of the present application, specifically, the clinic location significant aggregation representation vector and the communication embedding cascade vector are interactively fused with implicit feature conditions to obtain a clinic location-communication feature significant interactive fusion representation vector, which can be expressed by the formula: ;in, and are the significant aggregation representation vector of the medical visit location and the communication embedding cascade vector, respectively. To add by location, and are the weight matrix and bias vector respectively, for function, is the joint implicit feature vector of the medical location and communication context, for function, represents the medical treatment location-communication condition feature vector, The characteristic values ​​of each position in the vector representing the significant aggregation of the medical consultation positions are as follows, is the eigenvalue of each position in the communication embedding cascade vector, is the characteristic value of each position in the medical treatment location-communication condition characteristic vector, is the number of eigenvalues ​​in the significant aggregation representation vector of the medical visit location and the concatenated vector of the communication embedding, represents the logarithmic function value with base 2, represents an exponential function with the natural constant e as the base, is the semantic contribution of the clinic location, Contribution to communication semantics, and They are the normalized semantic contribution of the medical location and the normalized semantic contribution of communication, and They are the significant aggregation representation vector of the modulated visit position and the cascade embedding vector of the modulated communication, yes The transposed vector of yes Length, represents vector multiplication, yes function, It is the representation vector of the significant interaction fusion of the medical treatment location and communication features.

[0054] In step S134, based on the significant interaction fusion representation of the medical location-communication feature, the abnormal identification result is obtained, and the abnormal identification result is used to indicate whether the medical behavior of the target user is abnormal. Specifically, in the embodiment of the present application, based on the significant interaction fusion representation of the medical location-communication feature, the abnormal identification result is obtained, including: passing the medical location-communication feature significant interaction fusion representation vector through the medical behavior abnormality identifier based on the classifier to obtain the abnormal identification result, and the abnormal identification result is used to indicate whether the medical behavior of the target user is abnormal. That is, the medical location-communication feature significant interaction fusion representation of the medical location-communication feature that is conditionally guided to interact is used to perform classification processing, so as to intelligently determine whether the medical behavior of the target user is abnormal. In this way, the complex interactions between different features can be captured, not just simple linear relationships, so as to more accurately identify abnormal behaviors. At the same time, it can continuously optimize itself as the means of fraud evolve, so as to have better dynamic adaptability to cope with the ever-changing fraudulent behaviors, and improve the accuracy and intelligence of the review of insurance abnormal behaviors.

[0055] In particular, when the medical location significant aggregation representation vector and the communication embedding cascade vector respectively represent the embedded coding temporal aggregation features of the medical location information data set and the personal communication information data set, when conditional information is guided to significantly interact with them, the medical location-communication feature significant interaction fusion representation vector will also cause semantic interaction representation diversity due to the joint implicit conditional guidance differences under the embedding coding expressions of different data semantic features, thereby affecting the iterative consistency of the classification mapping and reducing the accuracy of the anomaly recognition results.

[0056] Preferably, the medical treatment location-communication feature significant interaction fusion representation vector is passed through a medical treatment behavior abnormality identifier based on a classifier to obtain the abnormal identification result, including: determining the medical treatment location-communication feature significant interaction fusion probability value corresponding to the medical treatment location-communication feature significant interaction fusion representation vector, the medical treatment location-communication feature significant interaction fusion probability value indicating the probability that the medical treatment behavior of the target user is abnormal; calculating the product of the first medical treatment location-communication feature significant interaction fusion feature value of the medical treatment location-communication feature significant interaction fusion representation vector and the medical treatment location-communication feature significant interaction fusion probability value to obtain the first a visit location-communication feature significant interaction fusion feature probability product; calculate the product of the second visit location-communication feature significant interaction fusion feature value of the visit location-communication feature significant interaction fusion representation vector and the difference between one minus the visit location-communication feature significant interaction fusion probability value to obtain the second visit location-communication feature significant interaction fusion feature inverse probability product; calculate the quotient of the first visit location-communication feature significant interaction fusion feature value divided by the difference between one minus the visit location-communication feature significant interaction fusion probability value to obtain the first visit location-communication feature significant interaction fusion feature inverse probability quotient; calculate the second visit location-communication feature significant interaction fusion feature value; calculate the product of .... The quotient of the location-communication feature significant interaction fusion characteristic value and the said medical location-communication feature significant interaction fusion probability value is obtained to obtain the second medical location-communication feature significant interaction fusion characteristic probability quotient; after adding the first medical location-communication feature significant interaction fusion characteristic probability product and the second medical location-communication feature significant interaction fusion characteristic inverse probability product, the difference between the first medical location-communication feature significant interaction fusion characteristic inverse probability quotient and the second medical location-communication feature significant interaction fusion characteristic probability quotient is subtracted to obtain the medical location-communication feature significant interaction fusion correction characteristic value; for the said medical location-communication feature For each first visit location - communication feature significant interaction fusion feature value and second visit location - communication feature significant interaction fusion feature value of the feature significant interaction fusion representation vector, matrix multiplication of the visit location - communication feature significant interaction fusion correction feature matrix composed of the visit location - communication feature significant interaction fusion correction feature values ​​is performed with the visit location - communication feature significant interaction fusion representation vector to obtain the corrected visit location - communication feature significant interaction fusion representation vector; the corrected visit location - communication feature significant interaction fusion representation vector is passed through the classifier-based visit behavior anomaly identifier to obtain the anomaly identification result.

[0057] The correction process of the medical location-communication feature significant interaction fusion representation vector is expressed as: ;in, is the significant interaction fusion representation vector of the medical location-communication feature, is the first The first medical location-communication feature significant interaction fusion feature value of the location, is the first The second medical location-communication feature of the location has a significant interactive fusion feature value, is the probability that the target user's medical treatment behavior is abnormal, The first The significant interaction fusion correction eigenvalues ​​of the medical location-communication characteristics of each location, Correction feature matrix for the significant interaction fusion of the visit location-communication features represents matrix multiplication, A representation vector for the corrected visit location-communication feature significant interaction fusion.

[0058] That is, the present application targets the region-boundary integral relationship of the high-dimensional feature manifold of the medical consultation location-communication feature significant interaction fusion representation vector, and approximates the simply connected region representation of the high-dimensional manifold of the feature set of the medical consultation location-communication feature significant interaction fusion representation vector composed of eigenvalue pairs by defining the overall probability distribution of the feature set of the medical consultation location-communication feature significant interaction fusion representation vector as the manifold constraint boundary, so as to avoid the ambiguity in the mapping of the diversified feature representation of the medical consultation location-communication feature significant interaction fusion representation vector to the local manifold representation in the high-dimensional feature class convergence space, improve the iterative consistency of the execution of each local feature distribution in the mapping task, so as to improve the feature classification convergence effect, and improve the accuracy of the abnormal recognition result obtained by the medical consultation behavior abnormality identifier based on the classifier of the medical consultation location-communication feature significant interaction fusion representation vector.

[0059] In summary, the deep learning-based insurance abnormal behavior identification method based on the embodiment of the present application is explained, which uses deep learning-based data processing technology to embed the data set of personal communication information and then cascade it in the entire time domain, and embed the data set of medical location information and perform significant aggregation between feature effects, so as to intelligently judge whether the medical behavior of the target user is abnormal based on the conditional guidance of significant interaction representation between the communication embedded cascade feature and the significant aggregation representation of the medical location. In this way, the complex interactions between different features can be captured, so as to more accurately identify abnormal behaviors. At the same time, it can continuously optimize itself as the means of fraud evolve, so as to have better dynamic adaptability to cope with the ever-changing fraudulent behaviors, and improve the accuracy and intelligence of the review of insurance abnormal behaviors.

[0060] Figure 6: is a system block diagram of an insurance abnormal behavior identification system based on deep learning according to an embodiment of the present application. Figure 6 As shown, according to the embodiment of the present application, the insurance abnormal behavior identification system 100 based on deep learning includes: an insurance audit information receiving module 110, which is used to receive the insurance audit information sent by the insurance institution; a personal communication and medical location information extraction module 120, which is used to extract the target user's personal communication information data set and the medical location information data set from the insurance audit information; an insurance audit information abnormality identification module 130, which is used to perform abnormal identification on the medical behavior of the target user based on the personal communication information data set and the medical location information data set to obtain the abnormal identification result of the insurance audit information, wherein the insurance audit information abnormality identification module includes: a personal communication and medical location information embedding coding unit, which is used to perform abnormal identification on the personal communication information data set and the medical location information data set The invention discloses a method for generating a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, which is used to generate a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, and a medical visit location-communication feature-based medical treatment behavior abnormality identification unit. The method comprises the following steps: performing a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, and a medical visit location-communication feature-based medical treatment behavior abnormality identification unit. The method comprises the following steps: performing a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, a medical visit location-communication feature-based medical treatment behavior abnormality identification unit, and a medical visit location-communication feature-based medical treatment behavior abnormality identification unit. The medical visit location-communication feature-based medical treatment behavior abnormality identification unit is used to obtain the medical visit behavior abnormality of the target user based on the medical visit location-communication feature-based medical treatment behavior abnormality identification result. The medical visit behavior abnormality identification result is used to indicate whether the medical visit behavior of the target user is abnormal.

[0061] Here, those skilled in the art can understand that the specific functions and operations of each unit and module in the above-mentioned insurance abnormal behavior identification system 100 based on deep learning have been referred to above. Figures 1 to 5 The description of the deep learning-based insurance abnormal behavior identification method has been introduced in detail, and therefore, its repeated description will be omitted.

[0062] In summary, the deep learning-based insurance abnormal behavior identification system 100 based on the embodiment of the present application is explained, which uses deep learning-based data processing technology to embed the data set of personal communication information and then cascade it in the entire time domain, and embed the data set of medical location information and perform significant aggregation between feature effects, so as to intelligently judge whether the medical behavior of the target user is abnormal based on the conditional guidance of significant interaction representation between the communication embedded cascade feature and the significant aggregation representation of the medical location. In this way, the complex interactions between different features can be captured, so as to more accurately identify abnormal behaviors. At the same time, it can continuously optimize itself as the means of fraud evolve, so as to have better dynamic adaptability to cope with the ever-changing fraudulent behaviors, and improve the accuracy and intelligence of the review of insurance abnormal behaviors.

Claims

1. A method for identifying abnormal insurance behavior based on deep learning, characterized in that: include: Receive insurance review information sent by insurance institutions; Extracting a data set of personal communication information and a data set of medical treatment location information of the target user from the insurance review information; Based on the data set of the personal communication information and the data set of the medical treatment location information, anomaly identification is performed on the medical treatment behavior of the target user to obtain an anomaly identification result of the insurance review information, including: embedding coding is performed on the data set of the personal communication information and the data set of the medical treatment location information to obtain a time series of communication embedded coding features and a time series of medical treatment location embedded coding features; significant aggregation of inter-feature forces is performed on the time series of medical treatment location embedded coding features to obtain a significant aggregation representation of medical treatment locations; The time series of the communication embedded coding features is cascaded to obtain the communication embedded cascade features, and then the conditional information is guided to significantly interact with the significant aggregation representation of the medical location to obtain the significant interaction fusion representation of the medical location-communication features; based on the significant interaction fusion representation of the medical location-communication features, the anomaly recognition result is obtained, and the anomaly recognition result is used to indicate whether there is an abnormality in the medical behavior of the target user.

2. The insurance abnormal behavior identification method based on deep learning according to claim 1 is characterized in that: Embedding and encoding the dataset of the personal communication information and the dataset of the medical visit location information to obtain a time series of communication embedded coding features and a time series of medical visit location embedded coding features, including: embedding and encoding the dataset of the personal communication information through a communication embedding matrix to obtain a time series of communication embedded coding vectors as the time series of the communication embedded coding features; embedding and encoding the dataset of the medical visit location information through a medical visit location embedding matrix to obtain a time series of medical visit location embedded coding vectors as the time series of the medical visit location embedded coding features.

3. The method for identifying abnormal insurance behavior based on deep learning according to claim 2 is characterized in that: The time series of the embedded coding features of the visit locations is subjected to significant aggregation of the forces between the features to obtain a significantly aggregated representation of the visit locations, including: calculating the static energy factor of each visit location embedded coding vector in the time series of the visit location embedded coding vector to obtain a sequence of the visit location static energy factors; selecting the visit location embedded coding vector corresponding to the largest visit location static energy factor in the sequence of the visit location static energy factors as the initial center vector of the visit location clustering; based on the sequence of the visit location static energy factors and the visit location clustering initial center vector, performing energy distribution-based feature aggregation on the time series of the visit location embedded coding vector to obtain a significantly aggregated representation vector of the visit locations as the significantly aggregated representation of the visit locations.

4. The method for identifying abnormal insurance behavior based on deep learning according to claim 3 is characterized in that: Calculate the static energy factor of each visit position embedded coding vector in the time series of the visit position embedded coding vector to obtain a sequence of visit position static energy factors, including: calculating the mean and variance of the visit position embedded coding vector to obtain the visit position mean and the visit position variance; subtracting the visit position embedded coding vector from the visit position mean by position, and calculating the fourth power of each position of the subtracted eigenvector to obtain the visit position difference modulation vector; calculating the expected value of the visit position difference modulation vector to obtain the visit position expected value; dividing the visit position expected value by the square of the visit position variance, and inputting the divided value into a sigmoid function to obtain the visit position static energy factor.

5. The method for identifying abnormal insurance behavior based on deep learning according to claim 4 is characterized in that: Based on the sequence of static energy factors of the visit location and the initial center vector of the visit location cluster, the time series of the visit location embedded coding vectors is subjected to feature aggregation based on energy distribution to obtain a significantly aggregated representation vector of the visit location, including: based on the spatial span between each visit location embedded coding vector in the time series of the visit location embedded coding vector and the initial center vector of the visit location cluster and the static energy factor of each visit location embedded coding vector and the static energy factor of the initial center vector of the visit location cluster, calculating the dynamic aggregation energy factor of each visit location embedded coding vector in the time series of the visit location embedded coding vector to obtain a sequence of dynamic aggregation energy factors of the visit location; inputting the sequence of dynamic aggregation energy factors of the visit location into a gated mask unit to obtain a sequence of dynamic aggregation weight factors of the visit location; and calculating the weighted sum of the time series of the visit location embedded coding vectors using the sequence of dynamic aggregation weight factors of the visit location to obtain the significantly aggregated representation vector of the visit location.

6. The method for identifying abnormal insurance behavior based on deep learning according to claim 5, characterized in that: The time series of the communication embedding coding features are cascaded to obtain the communication embedding cascade features, and then the conditional information is guided significantly to interact with the visit location significant aggregation representation to obtain the visit location-communication feature significant interaction fusion representation, including: cascading the time series of the communication embedding coding vectors to obtain the communication embedding cascade vector as the communication embedding cascade features; conditional information is guided significantly to interact with the visit location significant aggregation representation vector and the communication embedding cascade vector to obtain the visit location-communication feature significant interaction fusion representation vector as the visit location-communication feature significant interaction fusion representation.

7. The method for identifying abnormal insurance behavior based on deep learning according to claim 6, characterized in that: The significantly aggregated representation vector of the medical consultation location and the communication embedding cascade vector are significantly interacted with each other under the guidance of conditional information to obtain a significantly interactive fusion representation vector of the medical consultation location-communication feature, including: inputting the significantly aggregated representation vector of the medical consultation location and the communication embedding cascade vector into a joint implicit feature capture network to obtain a joint implicit feature vector of the medical consultation location-communication context; performing feature activation based on the Sigmoid function on the joint implicit feature vector of the medical consultation location-communication context to obtain a medical consultation location-communication condition feature vector; calculating the medical consultation location semantic contribution of the significantly aggregated representation vector of the medical consultation location relative to the medical consultation location-communication condition feature vector; calculating the communication semantic contribution of the communication embedding cascade vector to the medical consultation location-communication condition feature vector; and activating the medical consultation location semantic contribution to the medical consultation location-communication condition feature vector. The contribution degree and the communication semantic contribution degree are normalized, and the normalized visit location semantic contribution degree and the normalized communication semantic contribution degree are used to perform weighted modulation on the visit location significant aggregation representation vector and the communication embedding cascade vector to obtain the modulated visit location significant aggregation representation vector and the modulated communication embedding cascade vector; using the modulated visit location significant aggregation representation vector as the query vector, the modulated communication embedding cascade vector as the key vector and the visit location-communication condition feature vector as the value vector, the modulated visit location significant aggregation representation vector, the modulated communication embedding cascade vector and the visit location-communication condition feature vector are input into a feature-to-feature significant guided interaction module based on a converter structure to obtain the visit location-communication feature significant interaction fusion representation vector.

8. The method for identifying abnormal insurance behavior based on deep learning according to claim 7, characterized in that: The significant aggregation representation vector of the medical location and the communication embedding cascade vector are input into a joint implicit feature capture network to obtain a joint implicit feature vector of the medical location-communication context, including: adding the significant aggregation representation vector of the medical location and the communication embedding cascade vector by position, multiplying the obtained medical location-communication sum vector with a weight matrix, and then adding it by position with a bias vector to obtain a joint interaction vector of the medical location-communication; and using a tanh function to process the joint interaction vector of the medical location-communication to obtain a joint implicit feature vector of the medical location-communication context.

9. The method for identifying abnormal insurance behavior based on deep learning according to claim 8, characterized in that: Based on the significant interaction fusion representation of the medical location-communication features, the abnormality recognition result is obtained, including: passing the medical location-communication features significant interaction fusion representation vector through a classifier-based medical behavior abnormality identifier to obtain the abnormality recognition result, and the abnormality recognition result is used to indicate whether the medical behavior of the target user is abnormal.

10. An insurance abnormal behavior recognition system based on deep learning, characterized in that: include: An insurance review information receiving module is used to receive insurance review information sent by an insurance institution; A personal communication and medical treatment location information extraction module is used to extract a data set of personal communication information and a data set of medical treatment location information of a target user from the insurance review information; The module for identifying anomalies in insurance audit information is used to identify anomalies in the medical treatment behavior of the target user based on the data set of the personal communication information and the data set of the medical treatment location information to obtain an anomaly identification result of the insurance audit information, wherein the module for identifying anomalies in insurance audit information includes: a personal communication medical treatment location information embedding coding unit, used to embed and encode the data set of the personal communication information and the data set of the medical treatment location information to obtain a time series of communication embedded coding features and a time series of medical treatment location embedded coding features; a medical treatment location significant aggregation representation generation unit, used to perform significant aggregation of the force between features on the time series of the medical treatment location embedded coding features to obtain a medical treatment location significant aggregation representation; a medical treatment location-communication feature significant interaction fusion representation generation unit, used to cascade the time series of the communication embedded coding features to obtain the communication embedded cascade feature, and then perform conditional information-guided significant interaction with the medical treatment location significant aggregation representation to obtain a medical treatment location-communication feature significant interaction fusion representation; a medical treatment behavior anomaly identification unit, used to obtain the anomaly identification result based on the medical treatment location-communication feature significant interaction fusion representation, and the anomaly identification result is used to indicate whether the medical treatment behavior of the target user is abnormal.

Images