Classification method, device, equipment and storage medium based on graph attention network

Abstract

The present invention discloses a classification method, device, equipment and storage medium based on a graph attention network, which can be widely used in the field of artificial intelligence technology; the method of the present invention comprises: obtaining user behavior data to be processed; performing feature processing on the user behavior data to obtain correlation factor data; performing feature cleaning processing on the user behavior data and the correlation factor data to obtain target factor data; performing numerical processing on the target factor data to obtain target feature data corresponding to the target factor data; inputting the target feature data into a preset graph attention network classification model for classification processing to obtain classified data, wherein the classified data includes abnormal behavior data and normal behavior data. The present invention can effectively improve the accuracy of predicting individual illegal risk behaviors.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a classification method, device, equipment and storage medium based on a graph attention network. Background Art

[0002] With the rapid development of artificial intelligence (AI), the identification of individual illegal risk behaviors has gradually become intelligent and freed from manual work. In related technologies, the individual illegal risk identification model used is usually based on historical illegal risk behavior data and specific rules for judging illegal risk behaviors formed by insurance companies based on experience. The data used to detect illegal risk behaviors is often limited to the claims data collected by insurance companies and a small amount of external data. Therefore, predicting individual illegal risk behaviors through the above data is prone to large prediction errors. Summary of the invention

[0003] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0004] The embodiments of the present invention provide a classification method, apparatus, device and storage medium based on a graph attention network, which can effectively improve the accuracy of predicting individual illegal risk behaviors.

[0005] In a first aspect, an embodiment of the present invention provides a classification method based on a graph attention network, comprising:

[0006] Obtain user behavior data to be processed;

[0007] Performing feature processing on the user behavior data to obtain correlation factor data;

[0008] Performing feature cleaning processing on the user behavior data and the associated factor data to obtain target factor data;

[0009] Numerical processing is performed on the target factor data to obtain target feature data corresponding to the target factor data;

[0010] The target feature data is input into a preset graph attention network classification model for classification processing to obtain classification data, wherein the classification data includes abnormal behavior data and normal behavior data.

[0011] In a second aspect, an embodiment of the present invention provides a classification device based on a graph attention network, comprising:

[0012] A data acquisition module, used to acquire user behavior data to be processed;

[0013] A processing module, used to perform feature processing on the user behavior data to obtain correlation factor data;

[0014] A feature cleaning processing module is used to perform feature cleaning processing on the user behavior data and the associated factor data to obtain target factor data;

[0015] A numerical processing module, used for performing numerical processing on the target factor data to obtain target feature data corresponding to the target factor data;

[0016] A data classification module is used to input the target feature data into a preset graph attention network classification model for classification processing to obtain classified data, wherein the classified data includes abnormal behavior data and normal behavior data.

[0017] In the third aspect, an embodiment of the present invention provides a classification device based on a graph attention network, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the classification method based on the graph attention network of the previous embodiment is implemented.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing a computer-executable program, wherein the computer-executable program is used to execute the classification method based on the graph attention network of the previous embodiment.

[0019] The beneficial effects of the embodiments of the present invention include: first obtaining the user behavior data to be processed; then performing feature processing on the user behavior data to obtain the associated factor data; performing feature cleaning on the user behavior data and the associated factor data to obtain the target factor data; then performing numerical processing on the target factor data to obtain the target feature data corresponding to the target factor data; and finally inputting the target feature data into a preset graph attention network classification model for classification processing to obtain classification data, wherein the classification data includes abnormal behavior data and normal behavior data. In this embodiment, the target feature data obtained after feature processing, feature cleaning, and numerical processing on the user behavior data to be processed are more representative, and then the target feature data are classified and processed by the graph attention network classification model, and finally abnormal behavior data and normal behavior data can be obtained, and the abnormal behavior data indicates that there is a high probability of individual illegal risk behavior / individual illegal risk behavior. Compared with the related art, the embodiments of the present invention can effectively improve the prediction accuracy of individual illegal risk behavior.

[0020] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the technical solution of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the technical solution of the present invention and do not constitute a limitation on the technical solution of the present invention.

[0022] Figure 1 Schematic diagram of a flow chart of a classification method based on a graph attention network according to an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of a process flow of correlation factor data according to an embodiment of the present invention;

[0024] Figure 3 A schematic diagram of a data association process according to an embodiment of the present invention;

[0025] Figure 4 A schematic diagram of a process flow of target factor data according to an embodiment of the present invention;

[0026] Figure 5 A schematic diagram of a numerical processing flow in an embodiment of the present invention;

[0027] Figure 6 Schematic diagram of the structure of a classification device based on a graph attention network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0030] With the rapid development of artificial intelligence, the identification of individual illegal risk behaviors has gradually become intelligent and freed from manual work. In related technologies, the individual illegal risk identification model used is usually based on historical illegal risk behavior data and specific rules for judging illegal risk behaviors formed by insurance companies based on experience. The data that can be used to detect illegal risk behaviors is often limited to the claims data collected by insurance companies and a small amount of external data. In addition, the modeling method for detecting illegal risk behaviors usually only identifies them through the prediction of a single individual's characteristics. Therefore, predicting individual illegal risk behaviors in the above manner is prone to large prediction errors.

[0031] Based on this, the embodiments of the present invention provide a classification method, device, equipment and storage medium based on a graph attention network. The embodiments of the present invention can effectively improve the accuracy of predicting individual illegal risk behaviors.

[0032] Specifically, the embodiment of the present invention is based on a graph neural network (GNN), and specifically adopts a graph attention network classification model. When predicting individual illegal risk behaviors, it not only takes into account the characteristics of the current individual, but also the characteristics of other individuals associated with the individual. For example, assuming that the current individual has no illegal risk behaviors, but when his family members have illegal risk behaviors, the risk of the individual engaging in illegal behaviors will also increase, so as to improve the accuracy of the prediction.

[0033] The embodiment of the present invention is based on a graph neural network, integrating relevant feature data (such as user behavior data) and relationship data (such as correlation factor data) of the insurance industry, and classifying the relevant feature data and relationship data through a constructed claims anti-illegal risk behavior model, namely, a graph attention network classification model. Compared with models that do not use relationship data, the embodiment of the present invention can achieve better prediction accuracy.

[0034] It is understandable that the embodiments of the present application can acquire and process relevant data such as user behavior data of the embodiments of the present invention based on artificial intelligence technology (for example, data processing can be performed on correlation factor data, target factor data, target feature data, etc.). Among them, artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0035] It is understandable that the basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, and machine learning / deep learning.

[0036] Specifically, the user behavior data to be processed can be obtained through a terminal / device, and the terminal / device can be a mobile terminal device or a non-mobile terminal device. Among them, the mobile terminal device can be a mobile phone, a tablet computer, a laptop computer, a handheld computer, a PDA, an ultra-mobile personal computer (UMPC), a wearable device, a netbook, a personal digital assistant (PDA), an augmented reality (AR) / virtual reality (VR) device, etc.; the non-mobile terminal device can be a personal computer, a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present invention.

[0037] Specifically, refer to Figure 1 , an embodiment of the present invention provides a classification method based on a graph attention network, including but not limited to the following steps:

[0038] Step S100: obtaining user behavior data to be processed;

[0039] Step S200: Perform feature processing on user behavior data to obtain correlation factor data;

[0040] Step S300: Perform feature cleaning on user behavior data and related factor data to obtain target factor data;

[0041] Step S400: numerically process the target factor data to obtain target feature data corresponding to the target factor data;

[0042] Step S500: input the target feature data into a preset graph attention network classification model for classification processing to obtain classified data, wherein the classified data includes abnormal behavior data and normal behavior data.

[0043] It is understandable that the embodiments of the present invention need to obtain the user behavior data to be processed. For example, multiple, for example, more than 200 user behavior data can be integrated based on the Spark (a fast and general computing engine designed for large-scale data processing) big data platform. In other embodiments, other amounts of user behavior data can also be obtained, and the embodiments of the present invention do not specifically limit this. The user behavior data to be processed in the embodiments of the present invention can be data related to auto insurance claims.

[0044] The selection of multiple user behavior data is performed through the Spark platform, that is, the underlying factors of the acquired original data are selected to integrate multiple types of multiple user behavior data. It can be understood that the user behavior data includes at least one of historical claim behavior factors, accident site behavior factors, historical repair behavior factors, insurance behavior factors, insured person behavior factors or historical time series behavior factors.

[0045] Specifically, the historical claim behavior factor may be a historical claim record, including a historical claim number factor, a historical claim amount factor, a historical claim frequency factor, etc.;

[0046] The behavior factor of the accident scene can record the accident scene, including whether there are people injured, the factor of the damaged part of the vehicle, the factor of the degree of vehicle loss, the weather factor of the accident scene, and the factor of whether there is illegal loading, etc.

[0047] Historical repair behavior factors can be repair records of repair shops, including rescue cost factors, repair days factors, vehicle repair part factors, vehicle repair cost factors, vehicle residual value factors, etc.

[0048] Insurance behavior factors can be insurance policy information records, including factors such as the number of inquiries before reporting, premium factors, vehicle usage type (e.g., commercial / home use), and insurance channel factors;

[0049] The insured person's behavior factor can be the insured person's information record, including gender factor, grade factor, annual income level factor, occupation type factor, historical insurance policy quantity factor, customer value level factor, etc.;

[0050] Historical time series behavior factors can be used to record time series feature information, including latitude and longitude factors, location factors, WiFi connection factors, etc.

[0051] After obtaining multiple user behavior data to be processed, feature processing is performed on the multiple user behavior data to obtain correlation factor data, and then feature cleaning processing is performed on the user behavior data and the correlation factor data after feature processing, that is, feature selection processing is performed to obtain target factor data, and the target factor data is numerically processed to obtain target feature data corresponding to the target factor data, and finally the target feature data is input into a preset graph attention network classification model for classification processing to obtain classified data.

[0052] It can be understood that user behavior data is data that can be integrated through the Spark platform by selecting underlying factors, such as repair records of repair shops, insurance policy information records, etc., while associated factor data is data that requires further feature processing of user behavior data, that is, relational data. When predicting individual illegal risk behaviors, the embodiments of the present invention not only consider the characteristics of the current individual (such as user behavior data), but also consider the characteristics of other individuals related to it (that is, associated factor data), thereby improving the prediction accuracy.

[0053] It can be understood that the classified data of the embodiment of the present invention includes abnormal behavior data and normal behavior data, wherein the abnormal behavior data can be used to characterize the presence of individual illegal risk behavior / the probability of individual illegal risk behavior is high, while the normal behavior data can be used to characterize the absence of individual illegal risk behavior / the probability of individual illegal risk behavior is low. The target feature data obtained by this embodiment after feature processing, feature cleaning, and digitization of the user behavior data to be processed is more representative, that is, the embodiment of the present invention not only takes into account the characteristics of the current individual, but also the characteristics of other individuals related to it. After that, the target feature data is classified and processed through the graph attention network classification model, and finally abnormal behavior data and normal behavior data can be obtained, and the abnormal behavior data characterizes the presence of individual illegal risk behavior / the probability of individual illegal risk behavior is high. Compared with the related art, the embodiment of the present invention can effectively improve the prediction accuracy of individual illegal risk behavior.

[0054] Reference Figure 2 , user behavior data includes historical time series behavior factors, and feature processing is performed on user behavior data to obtain correlation factor data, including but not limited to the following steps:

[0055] Step S201: Perform feature processing on historical time series behavior factors to obtain behavior trajectory data and historical non-whitelist data;

[0056] Step S202: Use the behavior trajectory data and the historical non-whitelist data as correlation factor data.

[0057] It is understandable that the embodiment of the present invention needs to perform feature processing on the user behavior data in order to obtain the correlation factor data. The feature processing of the embodiment of the present invention can be LBS (Location Based Services) factor processing, network factor processing, distance factor processing, etc.

[0058] In step S201, the behavior trajectory data and the historical non-whitelist data are obtained by processing the historical time series behavior factors. Specifically, the historical time series behavior factors include latitude and longitude factors, location factors and wifi connection factors.

[0059] For example, by performing feature processing on the latitude and longitude factors and the location factors, the behavior trajectory data can be obtained. For example, based on the latitude and longitude factors of the vehicle and the LBS factors related to the POI (Point of Interest), the life trajectory of the user within a preset time period is processed to obtain the behavior trajectory data, such as the frequency of going to bars, parks, restaurants, cafes, etc., the latest time period for returning home, etc. The behavior trajectory data is used to characterize the life trajectory of the user in the recent period of time.

[0060] It is understandable that current applications such as APP (Application) usually record the user's longitude and latitude information, but cannot obtain detailed information about the location of the longitude and latitude, such as a park or a company, etc., while POI (Point of Interest) can record longitude and latitude information and corresponding location information. Therefore, by taking longitude and latitude information as longitude and latitude factors, and corresponding location information as location factors, and aggregating longitude and latitude factors and location factors, behavior trajectory data can be obtained, such as where the user has been every day, how long he has stayed, how many times he has been to the bar every month, where his workplace is, what his possible occupation is, etc. Therefore, by performing feature processing on longitude and latitude factors and location factors, behavior trajectory data can be obtained, and the behavior trajectory data can be used as correlation factor data.

[0061] It is understandable that for POI processing, in the geographic information system, a POI can be a house, a shop, a mailbox, a bus stop, etc. Each POI contains multiple information, such as name information, category information, latitude and longitude information, classification information, etc. Comprehensive POI information is essential information for enriching navigation maps. Timely POI points of interest can remind users of detailed information about road branches and surrounding buildings, and can also facilitate navigation to find the places you need, and choose the most convenient and unobstructed road for route planning. Therefore, in the navigation map, POI directly affects the usability of navigation. The number of POI information points in the navigation map, the accuracy of the information, and the speed of information update will all affect the use of navigation.

[0062] It is understandable that the embodiment of the present invention can perform feature processing through scripts such as spark sql (a spark component used to process structured data, which provides a programmable abstract data model and can be regarded as a distributed SQL query engine).

[0063] For another example, by processing the wifi connection factor with features, historical non-whitelist data can be obtained. It can be understood that since the wifi connection factor corresponds to the user's device identification number, the user's historical non-whitelist data can be queried based on the wifi connection factor. For example, the historical non-whitelist data can indicate that the device identification number corresponding to the user has been recorded as a non-whitelist. Specifically, based on the wifi connection factor, the user's wifi connection information is obtained, and then the historical non-whitelist data can be obtained through feature processing. The historical non-whitelist data can be used as the correlation factor data.

[0064] According to the wifi connection factor, the historical non-whitelist data of the current user and the insurance fraud user can also be processed. For example, when the insurance fraud user has multiple mobile phones, since each mobile phone has a corresponding device identification number, after the insurance fraud user uses a mobile phone to perform non-compliant operations, the corresponding device identification number of the mobile phone will be recorded in the historical non-whitelist data. If the insurance fraud user resells / transfers the mobile phone to the current user, there is a connection between the insurance fraud user and the current user.

[0065] It is understandable that other user behavior data may also be processed for features to obtain correlation factor data. For example, the behavior factor at the accident site and the historical repair behavior factor are processed for features to obtain correlation factor data, and based on the correlation factor data, the distance between the accident site and the repair shop, whether the driver in the accident and the insured are the same person, etc., may be obtained. The correlation factor data obtained in the embodiment of the present invention is specifically data that may be associated with individual illegal risk behaviors.

[0066] It is understandable that the Spark platform can be used to perform feature processing on the longitude and latitude factors and location factors, and the latest time period for returning home and the frequency of stay at a location can be obtained by using group calculation. The feature processing of the wifi connection factor can also be based on the statistics of the connection frequency. If the frequency of the current user and abnormal users such as insurance fraud users connecting to the same wifi is greater than the preset threshold, it is considered to be highly correlated. The distance processing is based on the longitude and latitude between the two places, and the spherical distance calculation is used to obtain the distance between the two places.

[0067] Reference Figure 3 , input the target feature data into the preset graph attention network classification model for classification processing, including but not limited to the following steps:

[0068] Step S510, obtaining case report number data;

[0069] Step S520: performing data association processing on the case report number data and the corresponding target characteristic data to obtain first target association characteristic data;

[0070] Step S530: according to the preset relationship type data, the case report number data corresponding to the same preset relationship type data is used as the second target association feature data;

[0071] Step S540: input the first target-related feature data and the second target-related feature data into a preset graph attention network classification model for classification processing.

[0072] It is understandable that the embodiments of the present invention will further construct relational data through the acquired report number data, with the aim of associating the report number data with the corresponding target feature data (i.e., user behavior data and correlation factor data after feature cleaning and digitization) through certain relationships (i.e., data association processing).

[0073] It can be understood that the report number data is the unique identification number after the vehicle accident report. By constructing relationship data, the target feature data associated with the report number data, that is, the first target associated feature data, and the relationship between the report number data and the report number data, that is, the second target associated feature data, can be used as input data for the graph attention network classification model.

[0074] Specifically, for step S530, the preset relationship type data includes at least one of a wifi relationship, a device relationship, a mobile phone number relationship, a policy relationship, a LBS relationship, a repair shop relationship, and an ID number relationship. According to the preset relationship type data, the case report number data corresponding to the same preset relationship type data is used as the second target association feature data.

[0075] For example, for the same insurance policy number, if there have been two reports in history, then two report number data will be generated, indicating that a relationship has been established between the two report number data. They have the same preset relationship type data, such as ID number relationship, mobile phone number relationship, etc., and data association processing can be performed on them to obtain the second target association feature data.

[0076] In some embodiments, the second target associated characteristic data is stored in a data table format, and the second target associated characteristic data is a plurality of case report number data, and the plurality of case report number data are mapped and associated with each other through the same preset relationship type data.

[0077] Alternatively, for step S520, i.e., for the case number data, data association processing may be performed with the target feature data corresponding to the target factor data, for example, data association processing may be performed with the historical claim behavior factor and the historical repair behavior factor after feature cleaning and numerical processing, to obtain the first target association feature data. The first target association feature data may be stored in a data table, a document, or other format, and no specific limitation is made thereto. For example, the first target association feature data is in the form of a data table, the first column represents the case number data, and several columns after the first column represent the target feature data corresponding to the case number data, and the target feature data is the user behavior data and the association factor data after feature cleaning and numerical processing.

[0078] Reference Figure 4 , perform feature cleaning on the user behavior data and the associated factor data to obtain target factor data, including at least one of the following:

[0079] Step S310: Perform feature cleaning on the user behavior data and the associated factor data according to the preset saturation index to obtain target factor data; or

[0080] Step S320: Perform feature cleaning on the user behavior data and the associated factor data according to the preset correlation index to obtain the target factor data.

[0081] It is understandable that in order to make the final classified data more accurate, the embodiment of the present invention needs to perform feature selection, i.e., feature cleaning, on the user behavior data and the associated factor data that has been processed by feature processing to obtain the target factor data. It should be noted that the embodiment of the present invention can also perform feature cleaning on the user behavior data after obtaining the user behavior data to be processed to filter out redundant and / or dirty data (such as empty data volume), and then perform feature processing, numerical processing, data association processing and classification processing to further ensure the accuracy of the classified data.

[0082] The embodiment of the present invention performs feature cleaning processing based on a preset saturation index and a preset correlation index. Specifically, according to the preset saturation index, user behavior data and correlation factor data that do not meet the preset saturation index are eliminated to obtain target factor data. This is because user behavior data and correlation factor data that are lower than the preset saturation index are not helpful to the graph attention network classification model.

[0083] It can be understood that saturation is: the proportion of the amount of non-empty data of a certain user behavior data (correlation factor data) to the total amount of data of the user behavior data (correlation factor data). For example, saturation>50% is used as a preset saturation index to perform feature selection on user behavior data and correlation factor data. In other embodiments, it can also be other values, which are not specifically limited in the embodiments of the present invention.

[0084] For example, saturation can be non-empty data volume / full data volume. Assuming that a user behavior data (correlation factor data) corresponds to a gender factor, there are three value cases: male, female, and unknown. Among them, the data volume of "male" is 5, the data volume of "female" is 3, and the data volume of "unknown" is 2, then the saturation is (5+3) / (5+3+2)=0.8, that is, 80%, which meets the preset saturation index.

[0085] It is understandable that if the saturation is too low, then most of the user behavior data (correlation factor data) are "unknown", and thus cannot provide useful information and need to be eliminated.

[0086] Specifically, according to the preset correlation index, the user behavior data and related factor data that do not meet the preset correlation index can be eliminated to obtain the target factor data. This is because if the correlation with the graph attention network classification model is too high, it means that it basically matches the characteristics of a certain type, then there may be data leakage problems and it needs to be eliminated.

[0087] It can be understood that the correlation is the correlation coefficient between the value of the user behavior data (correlation factor data) and the value of the Y label output in the graph attention network classification model. If the absolute value of the correlation coefficient is too high, the user behavior data (correlation factor data) may be directly generated by the Y label, and the user behavior data (correlation factor data) needs to be eliminated.

[0088] For example, assume that the Y label value is: 1 / 0, where 1 represents that the user has individual illegal risk behavior / the probability of individual illegal risk behavior is high, that is, the prediction target required by the embodiment of the present invention. In the data preprocessing process, such as in the feature processing of user behavior data, a user behavior data x is processed by feature processing, and the user behavior data x is a score based on Y. For example, when Y=1, x=10; and when Y=0, x=0, that is, x is a function of Y x=f(Y). It can be seen that there is a high correlation between the user behavior data x and Y. However, in the actual environment, the Y label corresponding to the user behavior data x, that is, the prediction target, should be unknown, and the value corresponding to the user behavior data x should also be unknown. It is usually necessary to classify and predict through a classification model such as the graph attention network classification model of the embodiment of the present invention to obtain the correlation between the user behavior data x and Y. Therefore, the user behavior data (correlation factor data) indicates that there is a data leakage problem and needs to be eliminated in order to improve the prediction accuracy of the model.

[0089] It is understandable that the embodiment of the present invention may first perform feature cleaning processing on the user behavior data and the associated factor data according to a preset saturation index, and then perform feature cleaning processing again according to a preset correlation index to make the final target factor data more accurate.

[0090] Reference Figure 5 , numerically process the target factor data to obtain target feature data corresponding to the target factor data, including but not limited to at least one of the following:

[0091] Step S410: performing data normalization processing on the target factor data to obtain normalized target feature data corresponding to the target factor data; or,

[0092] Step S420: performing data binning processing on the target factor data to obtain binned target feature data corresponding to the target factor data; or,

[0093] Step S430: Perform discrete numerical processing on the target factor data to obtain discrete target feature data corresponding to the target factor data.

[0094] The embodiment of the present invention performs numerical processing on the target factor data, which may be: data normalization processing, data binning processing and discrete numerical processing. The above numerical processing may be performed separately or in combination. For example, step S410 is performed alone to obtain a continuous variable, or step S430 and step S420 are performed in combination to obtain a continuous variable, or step S430 is performed alone to obtain a discrete variable, or step S430, step S420 and step S410 are performed in combination to obtain a continuous variable. The embodiment of the present invention does not specifically limit this. The data normalization processing of the embodiment of the present invention can facilitate the stability of the model; the data binning processing can avoid overfitting of the model.

[0095] Discrete numerical processing may be: numerical processing may be performed on different target factor data respectively, for example, discrete features may be numerically converted through a One-hot encoding method or a Target Encoding conversion method, but is not limited to these two methods.

[0096] In the embodiment of the present invention, the target factor data is processed into discrete numerical values. For example, if the target factor data includes: historical claim times factor, vehicle loss degree factor, and rescue cost factor, then One-Hot encoding can be used, and the codes are [1, 0, 0], [1, 1, 0], and [1, 0, 1] in sequence. Alternatively, Target Encoding is used for conversion processing, that is, discrete numerical processing is performed. For example, the number of parameter values ​​contained in the historical claim times factor is 5, the number of parameter values ​​contained in the vehicle loss degree factor is 10, and the number of parameter values ​​contained in the rescue cost factor is 10. After the discrete numerical processing corresponding to Target Encoding is performed, the discrete target feature data corresponding to the historical claim times factor is 0.2, the discrete target feature data of the vehicle loss degree factor is 0.4, and the discrete target feature data of the rescue cost factor is 0.4.

[0097] In some embodiments, after the target factor data is discretely numerically processed, discrete target feature data corresponding to the target factor data is obtained, and then the values ​​corresponding to the discrete target feature data are binned. For example, the values ​​corresponding to the discrete target feature data are sorted, either from large to small or from small to large, and then divided into multiple groups of values ​​in equal proportion, and corresponding labels are set to obtain binned target feature data corresponding to the target factor data; then the values ​​corresponding to the binned target feature data are normalized to obtain normalized target feature data corresponding to the target factor data, that is, the value ranges of the value groups with different labels are converted into preset value ranges, i.e., the normalized target feature data.

[0098] For example, in an embodiment of the present invention, the discrete target feature data is binned, that is, sorted from small to large according to the corresponding numerical values, and then the top 10% of the discrete target feature data is set as group 1, and the Y label is marked as 1, the discrete target feature data ranked between 10% and 20% is set as group 2, and the Y label is marked as 2, ..., the discrete target feature data ranked between 90% and 100% is set as group 10, and the Y label is marked as 10, and then the binned target feature data after the data binning processing is normalized, so that the model generalization ability is stronger during the subsequent training of the graph attention network classification model.

[0099] It can be understood that the graph attention network classification model is the GAT model.

[0100] It is understood that the GAT model includes weight data, which is obtained by the following formula:

[0101] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0102]

[0103] Among them, e ij Characterizes the attention coefficient corresponding to the ij node, w ij Characterize the weight data, h i Characterizes the feature vector corresponding to the i-th node, h j represents the eigenvector corresponding to the jth node, and h i 、h j represents the first target associated feature data or the second target associated feature data, W represents the weight matrix, W∈R F ' ×F , F represents the dimension of the node feature vector corresponding to the input of the graph attention layer in the graph attention network classification model, F' represents the dimension of the new node feature vector corresponding to the output of the graph attention layer, and a represents the shared self-attention mechanism: R F ' ×F →R, T represents transposition, N i represents the set of neighboring nodes of node i, || represents the vector concatenation operator, LeakyReLU represents the nonlinear activation function, k is an integer and k∈N i , e ik Represents the attention coefficient corresponding to the ik node, and exp represents the exponential function with the natural constant e as the base.

[0104] It is understandable that the embodiment of the present invention adopts the GAT model, which is based on the Attention mechanism and can assign different weight data to different nodes. It relies on pairs of adjacent nodes during training, rather than on the specific network structure. GAT (Graph Attention Network) aggregates adjacent nodes through the self-attention mechanism, realizes adaptive matching of different weight data, and thus improves the accuracy of the model.

[0105] It can be understood that for a single graph attention layer, the input corresponding to the graph attention layer can be a set of feature vectors of a node, for example, the feature vector of node i h = {h1,h2,...,h i}, h i ∈R F ,i represents the number of nodes, h i represents the feature vector corresponding to the i-th node, and F represents the dimension of the node feature vector corresponding to the input of the graph attention layer. It can be understood that in some embodiments, h i It can be characterized as any first target associated characteristic data except the column where the case report number data is located.

[0106] After the graph attention layer in the GAT model of the embodiment of the present invention, a new feature vector is output. Assuming that the dimension of the new node feature vector corresponding to the output of the graph attention layer is F' (which can be any value and may not be equal to F), that is, F' is the dimension of the new node feature vector corresponding to this feature vector, then this node feature can be expressed as h'={h'1,h'2,...,h' i}, h' i ∈R F '. In the graph attention layer, through a weight matrix W, W∈R F ' ×F Acting on each node, and then using the self-attention mechanism to calculate an attention coefficient for each node, the e in the embodiment of the present invention ij Characterizes the attention coefficient corresponding to the ij node, e ik Represents the attention coefficient corresponding to the ik node, where e ik Can be used with e ij The same calculation formula principle is used to calculate the shared self-attention mechanism (self-attention) used in the embodiment of the present invention, represented by a, where a is an R F ' ×F →R; T represents the transpose, W∈R F ' ×Fis a weight matrix (h i It can be understood that the embodiment of the present invention allocates attention to the set of neighboring nodes of node i, that is, k∈N i For example, in the first target-related feature data (second target-related feature data), for node i, the adjacent node k of node i is the adjacent column or adjacent row of node i, h i Characterizes the first target-related feature data (second target-related feature data) corresponding to the i-th node.

[0107] Understandably, e ij represents the importance / influence coefficient of node i on node j (scalar), e ik Represents the importance / influence coefficient of node i on node k (scalar).

[0108] The embodiment of the present invention selects the parameter a∈R 2F 'A single-layer feedforward neural network, and LeakyReLU is used to perform nonlinear processing on a, specifically:

[0109] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0110] The embodiment of the present invention also uses softmax to normalize the adjacent nodes of the central node, thereby obtaining:

[0111]

[0112] Finally, the output feature data is obtained by weighting the features corresponding to the input data, ie, the input first target associated feature data and the input second target associated feature data.

[0113] It can be understood that for nodes i and j,<i,j> Data association processing can be performed through preset relationship type data, for example, second target association characteristic data can be obtained through step S530. For example, wifi relationship, repair shop relationship, device relationship, etc. in the preset relationship type data. Step S520 can also be performed, that is, through data association processing, to associate the case number data with the corresponding target characteristic data to obtain the first target association characteristic data.

[0114] Specifically, h i 、h jThe first target associated feature data or the second target associated feature data corresponding to node i and node j is usually represented by a feature vector. For example, in some embodiments, (historical claim number factor, historical claim amount factor, historical claim frequency factor, whether there is someone injured factor, vehicle damaged part factor, vehicle loss degree factor...) = [3, 1500, 0.33, 1, 3, 5...], then the feature vector [3, 1500, 0.33, 1, 3, 5...] of the embodiment of the present invention specifically represents: the value corresponding to the historical claim number factor is 3, the value corresponding to the historical claim amount factor is 1500, the value corresponding to the historical claim frequency factor is 0.33, the value corresponding to the whether there is someone injured factor is 1, the value corresponding to the vehicle damaged part factor is 3, the value corresponding to the vehicle loss degree factor is 5...

[0115] In addition, h i and h j A linear transformation is performed through the weight matrix W, and then the two eigenvectors are concatenated into one eigenvector through the || vector concatenation operator. Then the inner product operation is performed with a vector of the same dimension, i.e., a, and finally a scalar is calculated through the nonlinear activation function LeakyReLU, i.e., e ij , which is used to represent the attention coefficient corresponding to the ij node. After that, it is subjected to softmax normalization to obtain the final weight data w ij .

[0116] It is understandable that w ij express<i,j> The strength of the relationships between nodes.

[0117] Finally, the embodiment of the present invention obtains the output feature data corresponding to each node through weighted combination of features. For example, for the i-th node, the corresponding output feature data is obtained:

[0118]

[0119] After obtaining the output feature data, in order to obtain the classification data corresponding to the output feature data, the embodiment of the present invention will also use the preset NN function (correlation function) f(x) to map the output feature data corresponding to node i to the classification data, and obtain the Y label:

[0120] y i =f(h' i )

[0121] It is understandable that the f(x) function can be a simple single-layer neural network plus a sigmoid function, specifically:

[0122] f(h' i )=σ(Wh' i )

[0123] Therefore, the embodiment of the present invention obtains classified data by inputting the target feature data (the first target-related feature data and the second target-related feature data) into the preset graph attention network classification model for classification processing, and the classified data may be abnormal behavior data or normal behavior data. For example, in some embodiments, the Y label is 1, indicating abnormal behavior data, and the Y label is 0, indicating normal behavior data, which is not specifically limited in the embodiment of the present invention.

[0124] It is understandable that the embodiment of the present invention can also obtain target feature data corresponding to the target factor data through steps S100 to S400, and the target feature data can be used as training data to input the training data into the preset graph attention network classification model for training, thereby obtaining a trained graph attention network classification model. Afterwards, the first target-related feature data and the second target-related feature data to be classified are input into the trained graph attention network classification model for classification processing, thereby obtaining classified data.

[0125] The classification method based on the graph attention network of the embodiment of the present invention selects the underlying factors of the original data to integrate the user behavior data to be processed, and then performs feature processing (feature selection), feature cleaning, numerical processing and data association processing on the user behavior data, and then inputs the first target-related feature data and the second target-related feature data into the graph attention network classification model to obtain classified data, such as a data table with classification results. Among them, the abnormal behavior data determined by the graph attention network classification model indicates that it has individual illegal risk behavior / individual illegal risk behavior probability is high, that is, the possibility of illegal risk behavior of auto insurance is high.

[0126] The GAT model of the embodiment of the present invention can utilize the associated nodes<i,j> The invention uses big data and machine learning algorithms to empower insurance business, significantly improving the efficiency and accuracy of case handling by claims staff, and greatly improving the user's claims service experience, thereby reducing compensation losses and improving user satisfaction.

[0127] Reference Figure 6 , an embodiment of the present invention also provides a classification device based on a graph attention network, including but not limited to the following modules:

[0128] The data acquisition module 100 is used to acquire the user behavior data to be processed;

[0129] The processing module 200 is used to perform feature processing on the user behavior data to obtain correlation factor data;

[0130] The feature cleaning processing module 300 is used to perform feature cleaning processing on the associated factor data to obtain target factor data;

[0131] The numerical processing module 400 is used to perform numerical processing on the target factor data to obtain target feature data corresponding to the target factor data;

[0132] The data classification module 500 is used to input the target feature data into a preset graph attention network classification model for classification processing to obtain classified data, wherein the classified data includes abnormal behavior data and normal behavior data.

[0133] It should be noted that the contents of the method embodiments of the present invention are all applicable to the device embodiments. The functions specifically implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method, which will not be repeated here.

[0134] In addition, an embodiment of the present invention also provides a classification device based on a graph attention network, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0135] The processor and the memory may be connected via a bus or other means.

[0136] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0137] It should be noted that the classification device based on graph attention network in this embodiment can be applied as the classification method based on graph attention network in the above embodiment. The classification device based on graph attention network in this embodiment and the classification method based on graph attention network in the above embodiment have the same inventive concept, so these embodiments have the same implementation principles and technical effects, which will not be described in detail here.

[0138] The non-transient software program and instructions required to implement the classification method based on the graph attention network of the above embodiment are stored in the memory. When executed by the processor, the classification method based on the graph attention network in the above embodiment is executed, for example, the above described Figure 1 The method steps S100 to S500 in FIG. 2 , the method steps S201 to S202 in FIG. 2 , Figure 3 Steps S510 to S540 of the method, Figure 4 Steps S310 to S320 of the method, Figure 5 Method steps S410 to S430 in.

[0139] The above-described classification device embodiment based on the graph attention network is merely illustrative, wherein the units described as separate components may or may not be physically separated, i.e., may be located in one place, or may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0140] In addition, one embodiment of the present invention further provides a computer-readable storage medium, which stores a computer-executable program, and the computer-executable program is executed by a processor or a controller, for example, by a processor in the above-mentioned classification device embodiment based on the graph attention network, so that the above-mentioned processor can execute the classification method based on the graph attention network in the above-mentioned embodiment, for example, execute the above-mentioned Figure 1 The method steps S100 to S500, Figure 2 Steps S201 to S202, Figure 3 Steps S510 to S540 of the method, Figure 4 Method steps S310 to S320 in FIG. 5 , and method steps S410 to S430 in FIG. 5 .

[0141] It will be appreciated by those skilled in the art that all or some of the steps and systems in the disclosed method above may be implemented as software, firmware, hardware and appropriate combinations thereof. Some physical components or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or may be implemented as hardware, or may be implemented as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or a non-transitory medium) and a communication medium (or a temporary medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that may be used to store desired information and may be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0142] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned implementation mode. Technical personnel familiar with the field can also make various equivalent deformations or substitutions without violating the spirit of the present invention. These equivalent deformations or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A classification method based on graph attention network, characterized in that: include: Acquire user behavior data to be processed; wherein the user behavior data includes historical claim behavior factors, accident scene behavior factors, historical repair behavior factors, insurance behavior factors, insured behavior factors, and historical time series behavior factors; Performing feature processing on the user behavior data to obtain correlation factor data; Performing feature cleaning processing on the user behavior data and the associated factor data to obtain target factor data; Numerical processing is performed on the target factor data to obtain target feature data corresponding to the target factor data; Get the case number data; Performing data association processing on the case report number data and the corresponding target feature data to obtain first target association feature data; According to the preset relationship type data, the case report number data corresponding to the same preset relationship type data is used as the second target association feature data; Inputting the first target-related feature data and the second target-related feature data into a preset graph attention network classification model for classification processing to obtain classification data, wherein the classification data includes abnormal behavior data and normal behavior data; The step of performing feature processing on the user behavior data to obtain correlation factor data includes: Performing feature processing on the historical time series behavior factors to obtain behavior trajectory data and historical non-whitelist data; The behavior trajectory data and the historical non-whitelist data are used as the correlation factor data.

2. The classification method based on graph attention network according to claim 1, characterized in that: The performing feature cleaning processing on the user behavior data and the associated factor data to obtain target factor data includes at least one of the following: According to a preset saturation index, feature cleaning is performed on the user behavior data and the associated factor data to obtain the target factor data; or, According to the preset correlation index, the user behavior data and the correlation factor data are subjected to feature cleaning processing to obtain the target factor data.

3. The classification method based on graph attention network according to claim 1, characterized in that: The step of numerically processing the target factor data to obtain target feature data corresponding to the target factor data includes at least one of the following: Performing data normalization processing on the target factor data to obtain normalized target feature data corresponding to the target factor data; or, Performing data binning processing on the target factor data to obtain binned target feature data corresponding to the target factor data; or, The target factor data is processed into discrete numerical form to obtain discrete target feature data corresponding to the target factor data.

4. The classification method based on graph attention network according to claim 1, characterized in that: The graph attention network classification model is a GAT model, and the GAT model includes weight data, and the weight data is obtained by the following formula: e ij =LeakyReLU(a T [Wh i ||Wh j ]) Among them, the e ij Characterizes the attention coefficient corresponding to the ij node, the w ij Characterize the weight data, the h i Characterizes the feature vector corresponding to the i-th node, the h j represents the eigenvector corresponding to the jth node, and h i 、h j represents the first target-related feature data or the second target-related feature data, wherein W represents a weight matrix, W∈R F ' ×F , the F represents the dimension of the node feature vector corresponding to the input of the graph attention layer in the graph attention network classification model, F' represents the dimension of the new node feature vector corresponding to the output of the graph attention layer, and the a represents the shared self-attention mechanism: R F ' ×F →R, the T represents the transpose, the N i represents the set of adjacent nodes of node i, the || represents the vector concatenation operator, the LeakyReLU represents the nonlinear activation function, and k is an integer and k∈N i , the e ik Represents the attention coefficient corresponding to the ik node, and the exp represents an exponential function with the natural constant e as the base.

5. A classification device based on a graph attention network, characterized in that: include: A data acquisition module, used to acquire user behavior data to be processed; wherein the user behavior data includes historical claim behavior factors, accident scene behavior factors, historical repair behavior factors, insurance behavior factors, insured behavior factors, and historical time series behavior factors; A processing module, used to perform feature processing on the user behavior data to obtain correlation factor data; A feature cleaning processing module is used to perform feature cleaning processing on the user behavior data and the associated factor data to obtain target factor data; A numerical processing module, used for performing numerical processing on the target factor data to obtain target feature data corresponding to the target factor data; A data classification module is used to obtain case report number data; perform data association processing on the case report number data and the corresponding target feature data to obtain first target association feature data; according to preset relationship type data, use the case report number data corresponding to the same preset relationship type data as second target association feature data; input the first target association feature data and the second target association feature data into a preset graph attention network classification model for classification processing to obtain classification data, wherein the classification data includes abnormal behavior data and normal behavior data; The step of performing feature processing on the user behavior data to obtain correlation factor data includes: Performing feature processing on the historical time series behavior factors to obtain behavior trajectory data and historical non-whitelist data; The behavior trajectory data and the historical non-whitelist data are used as the correlation factor data.

6. A classification device based on a graph attention network, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the classification method based on the graph attention network as described in any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium, characterized in that: A computer executable program is stored, and the computer executable program is used to execute the classification method based on the graph attention network as described in any one of claims 1 to 4.

Images