Insurance fraud behavior detection method, device and equipment and readable storage medium
Through the combination of the knowledge graph embedding model and the link prediction model, entity combination processing and relational space embedding features, the problem of low detection accuracy of insurance fraud behavior is solved, and higher detection accuracy and comprehensiveness are achieved.
Patent Information
- Application Number
- CN202411182443.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-06
AI Technical Summary
The existing machine learning model-based insurance fraud detection accuracy is low, especially due to the small number of samples of fraud cases, which leads to inaccurate detection.
The knowledge graph embedding model and link prediction model are used to detect gang relationships in insurance cases through entity combination processing and relationship space embedding features to determine whether there is fraud.
The accuracy of insurance fraud detection is improved, and by tapping potential gang relationships, the inaccuracy problem of detection caused by small sample size is reduced, and more comprehensive and accurate detection results are provided.
Smart Images

Figure CN119090655A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and specifically to an insurance fraud detection method, device, computer equipment and computer-readable storage medium. Background Art
[0002] With the rapid development of social economy, criminal activities in the financial field have also increased dramatically. Among them, insurance fraud has had a huge impact on social economy and endangered the healthy development of economic construction. In related technologies, machine learning models are established to detect insurance cases with fraud. However, compared with the overall number of insurance cases, the number of insurance case samples with fraud is relatively small. Therefore, the detection accuracy of existing insurance case fraud based on machine learning models is relatively low. Summary of the invention
[0003] The present application provides an insurance fraud detection method, device, computer equipment and computer-readable storage medium, which belong to the field of artificial intelligence technology and can improve the accuracy of insurance fraud detection.
[0004] In a first aspect, the present application provides a method for detecting insurance fraud, the method comprising:
[0005] Acquire various important entities involved in the target insurance case and entity attributes of various important entities involved in the target insurance case;
[0006] According to a preset entity combination type, entity combination processing is performed on each important entity involved in the target insurance case to obtain an entity binary group of the target insurance case;
[0007] By using the trained knowledge graph embedding model, the relational space embedding features of the entity binary are obtained according to the entity attributes of each important entity involved in the target insurance case;
[0008] Using the trained link prediction model, group relationship detection is performed according to the relationship space embedding features of the entity binary group to obtain the group relationship detection result of the entity binary group;
[0009] Based on the gang relationship detection result of the entity binary group, determine whether there is fraud in the target insurance case.
[0010] In some embodiments, the trained knowledge graph embedding model is used to obtain the relational space embedding features of the entity binary according to the entity attributes of each important entity involved in the target insurance case, including:
[0011] By embedding the trained knowledge graph into the entity representation layer of the model, according to the entity attributes of each important entity involved in the target insurance case, each entity in the entity binary is represented by a vector, and the entity space embedding feature of the entity binary is obtained;
[0012] By embedding the relationship conversion layer in the trained knowledge graph model, the relationship space conversion is performed according to the entity space embedding features of the entity binary and the group relationship conversion matrix corresponding to the relationship conversion layer to obtain the relationship space embedding features of the entity binary.
[0013] In some embodiments, the group relationship detection is performed based on the relationship space embedding features of the entity binary group using the trained link prediction model to obtain the group relationship detection result of the entity binary group, including:
[0014] Using the trained link prediction model, prediction is performed based on the relational space embedding features of the entity binary to obtain the group relationship score of the entity binary;
[0015] Based on the group relationship score, a group detection result of the entity binary is determined.
[0016] In some embodiments, the method further comprises:
[0017] By using the trained entity representation layer in the knowledge graph embedding model, each entity in the entity binary is represented by a vector to obtain the entity space embedding feature of the entity binary, wherein the entity space embedding feature includes the entity space feature of the head entity and the entity space feature of the tail entity;
[0018] The trained link prediction model is used to predict the entity binary group according to the relationship space embedding feature of the entity binary group to obtain the group relationship score of the entity binary group, including:
[0019] The trained link prediction model is used to perform prediction based on the relational space embedding features of the entity bigrams and the entity space embedding features of the entity bigrams to obtain the group relationship score of the entity bigrams.
[0020] In some embodiments, the trained knowledge graph embedding model is obtained by the following steps:
[0021] Obtaining entity attributes of various important entities involved in the insurance case sample and entity relationships involved in the insurance case sample;
[0022] Constructing an insurance fraud knowledge graph based on entity attributes of various important entities involved in the insurance case samples and entity relationships involved in the insurance case samples;
[0023] According to the insurance fraud knowledge graph, a preset knowledge graph embedding model is trained to obtain a trained knowledge graph embedding model.
[0024] In some embodiments, the method further comprises:
[0025] Obtaining a first important entity from each important entity involved in the insurance case sample;
[0026] Using the trained knowledge graph embedding model, from each important entity involved in the insurance case sample, the second important entity with the highest probability of having a gang relationship with the first important entity is obtained;
[0027] The first important entity and the second important entity are taken as an entity pair having a gang relationship, and a preset link prediction model is trained to obtain a trained link prediction model.
[0028] In some embodiments, the method further comprises:
[0029] Acquire characteristic data of an insurance case sample, wherein the characteristic data includes a plurality of reference entities;
[0030] Training a preset fraud detection model according to the reference entity to obtain a trained fraud detection model;
[0031] Deleting the target reference entity in the feature data to obtain deleted feature data;
[0032] Using the deleted feature data to train a preset fraud detection model to obtain a retrained fraud detection model;
[0033] If the accuracy of the retrained fraud detection model is less than the accuracy of the trained fraud detection model, the target reference entity is taken as an important entity.
[0034] In a second aspect, the present application provides an insurance fraud detection device, the insurance fraud detection device comprising:
[0035] An acquisition unit, used to acquire various important entities involved in a target insurance case and entity attributes of various important entities involved in the target insurance case;
[0036] A combination unit, used to perform entity combination processing on each important entity involved in the target insurance case according to a preset entity combination type, to obtain an entity binary group of the target insurance case;
[0037] An embedding unit, used to obtain the relational space embedding features of the entity binary group according to the entity attributes of each important entity involved in the target insurance case through a trained knowledge graph embedding model;
[0038] A detection unit, configured to perform group relationship detection according to the relationship space embedding features of the entity binary group through a trained link prediction model, and obtain a group relationship detection result of the entity binary group;
[0039] The detection unit is further used to determine whether there is fraud in the target insurance case based on the gang relationship detection result of the entity binary group.
[0040] In a third aspect, the present application also provides a computer device, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the insurance fraud detection method when executing the computer program.
[0041] In a fourth aspect, the present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is loaded by a processor to execute the insurance fraud detection method.
[0042] In the present application, on the one hand, the relational space mapping of entity binary groups is performed through the trained knowledge graph embedding model to obtain the relational space embedding features of the entity binary groups, and the relational space embedding features are used to perform gang relationship detection to obtain the gang relationship detection results of the entity binary groups. The entities can be mapped from the entity space to the relational space, thereby exploring the gang relationships of potential collusion fraud, avoiding the problem of inaccurate fraud detection due to the small amount of sample data of insurance cases with fraudulent behavior, and improving the detection accuracy of insurance fraud behavior to a certain extent; on the other hand, entity binary groups are obtained by performing entity combination processing based on the various important entities involved in the target insurance case, and gang relationship detection can be performed on entity pairs of various combination types, and more entity pairs that may have fraud gang relationships in a single insurance case can be explored, thereby more comprehensively detecting whether there is a gang relationship in a single insurance case and improving the detection accuracy of insurance fraud behavior. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0044] Figure 1This is a flow chart of the insurance fraud detection method provided in the embodiment of the present application;
[0045] Figure 2 This is a flowchart of an embodiment of step 103 in the embodiment of the present application;
[0046] Figure 3 This is a flowchart of an embodiment of step 104 in the embodiment of the present application;
[0047] Figure 4 is another embodiment flow chart of step 104 in the embodiment of the present application;
[0048] Figure 5 This is a schematic diagram illustrating the overall process of an embodiment of the insurance fraud detection process in the embodiment of the present application;
[0049] Figure 6 It is a schematic diagram of the structure of an embodiment of an insurance fraud detection device provided in an embodiment of the present application;
[0050] Figure 7 It is a schematic block diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.
[0052] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.
[0053] In the description of the embodiments of the present application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.
[0054] In order to enable any person skilled in the art to implement and use the present application, the following description is provided. In the following description, details are listed for the purpose of explanation. It should be understood that those of ordinary skill in the art can recognize that the present application can also be implemented without using these specific details. In other examples, the known process will not be elaborated in detail to avoid unnecessary details that make the description of the present application embodiment obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the widest range of principles and features disclosed in accordance with the embodiments of the present application.
[0055] Embodiments of the present application provide a method, apparatus, computer device, and computer-readable storage medium for detecting insurance fraud.
[0056] The executor of the insurance fraud detection method in the embodiment of the present application may be the insurance fraud detection device provided in the embodiment of the present application, or the computer device provided in the embodiment of the present application, wherein the insurance fraud detection device may be implemented in hardware or software.
[0057] Some embodiments of the present application are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.
[0058] See also Figure 1 , Figure 1 1 is a flow chart of an insurance fraud detection method provided in an embodiment of the present application. The insurance fraud detection method includes steps 101 to 105, wherein:
[0059] 101. Obtain each important entity involved in a target insurance case and entity attributes of each important entity involved in the target insurance case.
[0060] Among them, the target insurance case refers to the insurance case that needs to be tested for fraud.
[0061] Among them, the important entities involved in insurance cases may be the policyholder, the insured, the vehicle, the institution, etc.
[0062] Among them, entity attributes are specific information or data used to describe the entity. For example, taking the entity as a vehicle, the entity attributes of the vehicle can be brand, model, color, manufacturing date, etc.
[0063] Since an insurance case contains many entities, in order to reduce the amount of entity data processing and improve the accuracy of fraud detection, in this embodiment, a fraud detection model is established to screen important entities. In addition, entities can be manually added, deleted or renamed, and the relationship between entities can be defined to improve the knowledge graph constructed in step B2. For example, important entities can be screened out through the following steps A1 to A5:
[0064] A1. Obtain characteristic data of insurance case samples, wherein the characteristic data includes multiple reference entities.
[0065] A2. Train a preset fraud detection model according to the reference entity to obtain a trained fraud detection model.
[0066] A3. Deleting the target reference entity in the feature data to obtain deleted feature data.
[0067] The target reference entity is the reference entity that needs to be deleted.
[0068] A4. Using the deleted feature data to train a preset fraud detection model to obtain a retrained fraud detection model.
[0069] A5. If the accuracy of the retrained fraud detection model is less than the accuracy of the trained fraud detection model, the target reference entity is taken as an important entity.
[0070] For example, in this embodiment, a preset fraud detection model is constructed using XGBoost (eXtreme Gradient Boosting, a gradient boosting decision tree), lightGBM (Light Gradient Boosting Machine, a gradient boosting decision tree) or a neural network. First, the preset fraud detection model is trained using the characteristic data of the insurance case sample to obtain a trained fraud detection model. Then, a reference entity is deleted each time, and then the preset fraud detection model is trained using the characteristic data after deleting the target reference entity to obtain a retrained fraud detection model. Compared with the trained fraud detection model, if the model performance of the retrained fraud detection model is significantly reduced (for example, the accuracy of the retrained fraud detection model is less than the accuracy of the trained fraud detection model, and the difference between the accuracy of the retrained fraud detection model and the accuracy of the trained fraud detection model is less than the threshold difference threshold), it indicates that the removed target reference entity plays an important role in the machine learning model of fraud detection. Therefore, the target reference entity can be used as an important entity. The knowledge graph constructed using these important entities in the future can improve the quality of the entity, filter out irrelevant entity data, reduce the amount of entity data processing, and improve the accuracy of fraud detection.
[0071] 102. Perform entity combination processing on each important entity involved in the target insurance case according to a preset entity combination type to obtain an entity tuple of the target insurance case.
[0072] In order to better understand this embodiment, some names involved in this embodiment are introduced below:
[0073] 1. Entity tuple (e1, e2): An entity tuple is an entity pair consisting of a head entity e1 and a tail entity e2. For example, the insured and the insured can be regarded as an entity tuple.
[0074] 2. Entity pair: Any two entities can form an entity pair. For example, the policyholder and the insured can be an entity pair.
[0075] Exemplarily, each important entity involved in the target insurance case can be entity combined according to the preset entity combination type to obtain entity tuples of various entity combination types. For example, the target insurance case involves 4 entities: policyholder A, insured B, vehicle C, and organization D, and 6 entity combination types are preset (combination types 1, 2, 3, 4, 5, and 6 respectively): combination type 1 is [policyholder, insured], combination type 2 is [insured, vehicle], combination type 3 is [policyholder, vehicle], combination type 4 is [policyholder, organization], combination type 5 is [insured, organization], and combination type 6 is [vehicle, organization], then 6 entity tuples can be combined, [policyholder A, insured B], [insured B, vehicle C], [policyholder A, vehicle C], [policyholder A, organization D], [insured B, organization D], and [vehicle C, organization D].
[0076] The number of entity tuples obtained in step 102 is set according to the actual business scenario requirements. For example, in step 102, only entity tuples of one entity combination type may be obtained. In step 104, when the entity tuples of one entity combination type exist in a gang relationship, it is determined that fraud exists in the target insurance case. In step 102, entity tuples of multiple entity combination types may also be obtained. In step 104, when the entity tuples of multiple entity combination types exist in a gang relationship, it is determined that fraud exists in the target insurance case.
[0077] 103. By using the trained knowledge graph embedding model, the relational space embedding features of the entity binary are obtained according to the entity attributes of each important entity involved in the target insurance case.
[0078] For ease of understanding, the knowledge graph embedding model involved in this embodiment is first introduced below. The knowledge graph embedding model includes an entity representation layer, a relationship representation layer, and a relationship conversion layer. The functions of each layer are as follows:
[0079] 1. Entity representation layer, used to project each entity into the entity space; specifically, embed each entity according to its entity attributes, map each entity into a vector of fixed length, and obtain the entity space features of each entity.
[0080] 2. The relationship representation layer is used to embed each relationship (i.e., the group relationship of each entity combination type), represent each relationship (i.e., the group relationship of each entity combination type) itself as a vector, and obtain the relationship vector of each relationship (i.e., the group relationship of each entity combination type). For example, the relationship representation layer can embed the group relationship 1 of entity combination type 1 (such as [policyholder, insured person]) and the group relationship 2 of entity combination type 2 (such as [policyholder, vehicle]), and obtain the relationship vector of group relationship 1 and the relationship vector of group relationship 2.
[0081] 3. Relationship conversion layer, which is used to project each entity in the entity pair from the entity space to the relationship space to obtain the relationship space features of each entity; specifically, for each relationship (i.e., the group relationship of each entity combination type), the mapping matrix Mr is used to project the entity from the entity space to the relationship space. For example, the corresponding projection vectors of the head entity and the tail entity are: hr=hMr, tr=tMr, where h and t are the entity space embedding representations of the head entity and the tail entity, respectively, and hr and tr are the relationship space embedding representations of the head entity and the tail entity, respectively.
[0082] Exemplarily, the knowledge graph embedding model can be trained through the following steps B1 to B3:
[0083] B1. Obtain entity attributes of each important entity involved in the insurance case sample, as well as entity relationships involved in the insurance case sample.
[0084] B2. Construct an insurance fraud knowledge graph based on the entity attributes of each important entity involved in the insurance case sample and the entity relationships involved in the insurance case sample.
[0085] Among them, the entity relationship includes the gang relationship of fraud. In this embodiment, if the frequency (n pb ) is greater than a preset threshold, it is considered that a fraudulent cooperation relationship (i.e., a fraudulent gang relationship) exists between the entity pairs of this entity combination type.
[0086] Among them, the knowledge graph is a structured semantic network used to represent and store entities, concepts, relationships, and the semantic associations between them. It can be seen as a semantic network that describes information in the real world through nodes (entities and concepts) and edges (relationships).
[0087] Exemplarily, triples can be formed according to "entity-relationship-entity", and an insurance fraud knowledge graph can be constructed based on the entity attributes of each important entity involved in the insurance case sample and the entity relationships involved in the insurance case sample.
[0088] B3. According to the insurance fraud knowledge graph, a preset knowledge graph embedding model is trained to obtain a trained knowledge graph embedding model.
[0089] Exemplarily, step B3 may be specifically as follows: 1) perform input data preparation such as knowledge graph representation and triple extraction. When representing the knowledge graph, the entities and relationships in the insurance fraud knowledge graph are represented in numerical form, such as using a unique ID or vector to represent each entity and relationship; and extract triples (head entity, relationship, tail entity) from the insurance fraud knowledge graph. 2) Use TransR as the preset knowledge graph embedding model. 3) Consider the gap between the score of the correct triple and the score of the wrong triple, ensure that the score of the correct triple is higher than the score of the wrong triple, and establish the loss function of the knowledge graph embedding model. 4) Negative sampling: For each positive example (correct triple), generate several negative examples (wrong triples) to form a training sample. 5) With the goal of minimizing the loss function, the gap between the score of the correct triple and the score of the wrong triple is as large as possible, and adjust the model parameters of the preset knowledge graph embedding model (including the parameters of the entity representation layer, the relationship representation layer, and the relationship conversion layer) according to the training sample, so as to obtain a trained knowledge graph embedding model.
[0090] Furthermore, since there are few samples of insurance cases with fraudulent behavior, it is difficult to fully discover entity pairs with fraud gang relationships. In order to increase the amount of entity pairs with fraud gang relationships and improve the accuracy of the link prediction model, in this embodiment, the trained knowledge graph embedding model is used to mine potential entity pairs with fraud gang relationships from insurance case samples, and the newly mined entity pairs with fraud gang relationships are used to train the link prediction model. Specifically, the following steps C1 to C3 can be performed:
[0091] C1. Obtaining a first important entity from various important entities involved in the insurance case sample.
[0092] Exemplarily, the first important entity may be any one of the important entities involved in the insurance case sample.
[0093] C2. Using the trained knowledge graph embedding model, the second important entity with the highest probability of having a gang relationship with the first important entity is obtained from various important entities involved in the insurance case sample.
[0094] The process of obtaining the second most important entity can be specifically as follows:
[0095] (1.1) Traverse each important entity involved in the insurance case sample, and use the current traversed entity and the first important entity as candidate entity pairs (e1, e2).
[0096] (1.2) Through the trained knowledge graph embedding model’s relational representation layer, the relational embedding representation is performed according to the entity combination type of the candidate entity pair (e1, e2) to obtain the relational vector r of the candidate entity pair (e1, e2).
[0097] (1.3) Through the trained knowledge graph embedding model’s entity representation layer, the candidate entity pair (e1, e2) is represented by entity embedding, and the entity space embedding features of the candidate entity pair (e1, e2) are obtained.
[0098] (1.4) Through the trained knowledge graph embedding model, the relational space transformation is performed according to the entity space embedding features of the candidate entity pair (e1, e2) to obtain the relational space embedding features (h, t) of the candidate entity pair (e1, e2).
[0099] (1.5) The relational space embedding features (h, t) of the candidate entity pair (e1, e2) and the relational vector of the candidate entity pair (e1, e2) are combined as the reconstructed triplet (h, r, t) of the current traversed entity.
[0100] (1.6) The reconstructed triplet (h, r, t) of the current traversal entity is scored using the preset triplet scoring function to obtain the triplet score of the current traversal entity.
[0101] (1.7) From the various important entities involved in the insurance case sample, obtain the important entity with the highest corresponding triple score.
[0102] (1.8) If the triple score of the important entity with the highest corresponding triple score is greater than the preset score threshold, the important entity with the highest corresponding triple score among the important entities involved in the insurance case sample is taken as the second important entity.
[0103] C3. Taking the first important entity and the second important entity as an entity pair having a gang relationship, training a preset link prediction model to obtain a trained link prediction model.
[0104] Specifically, the first important entity and the second important entity are constructed into an entity pair sample, and the entity pair sample is marked as having a group relationship, thereby obtaining the group relationship marking result of the entity pair sample; the entity pair sample is embedded and represented by the trained knowledge graph embedding model to obtain the relational space embedding features of the entity pair sample; then, a preset link prediction model is used to predict the group relationship prediction result of the entity pair sample based on the relational space embedding features of the entity pair sample; according to the loss between the group relationship prediction result of the entity pair sample and the group relationship marking result of the entity pair sample, the parameters of the preset link prediction model are adjusted to finally obtain the trained connection prediction model.
[0105] Among them, the specific implementation of "predicting the entity-pair group relationship prediction results based on the relationship space embedding features of the sample through the preset link prediction model" is similar to the implementation of step 104. For details, please refer to the relevant instructions of step 104, which will not be repeated here.
[0106] Therefore, through steps C1 to C3, the knowledge graph embedding model can be used to perform relationship inference, and potential entity pairs with fraud gang relationships can be mined from insurance case samples. Entity pair samples with gang relationships are constructed, and the link prediction model is trained. This can increase the amount of entity pair data with fraud gang relationships to improve the accuracy of the link prediction model.
[0107] For example, reference Figure 2 At this time, step 103 may specifically include the following steps 1031 to 1032:
[0108] 1031. By embedding the trained knowledge graph into the entity representation layer of the model, a vector representation is performed on each entity in the entity binary group according to the entity attributes of each important entity involved in the target insurance case, so as to obtain the entity space embedding features of the entity binary group.
[0109] The entity space embedding features of the entity bigram include the entity space features of the head entity in the entity bigram and the entity space features of the tail entity in the entity bigram.
[0110] 1032. By embedding the relationship conversion layer in the trained knowledge graph model, the relationship space conversion is performed according to the entity space embedding features of the entity binary and the group relationship conversion matrix corresponding to the relationship conversion layer to obtain the relationship space embedding features of the entity binary.
[0111] The relational space embedding features of the entity binary include the relational space features of the head entity in the entity binary and the relational space features of the tail entity in the entity binary.
[0112] Exemplarily, for each entity binary obtained by the combined processing in step 102, the relational space embedding process of steps 1031 to 103 is performed on each entity binary, respectively. Specifically, for each entity binary, in step 1031, the entity representation layer in the trained knowledge graph embedding model is used to represent the head entity and the tail entity in the entity binary according to the entity attributes of the head entity and the entity attributes of the tail entity in the entity binary, respectively, to obtain the entity space features of the head entity in the entity binary and the entity space features of the tail entity in the entity binary; then, in step 1032, the relational conversion layer in the trained knowledge graph embedding model is used to represent the head entity and the tail entity in the entity binary according to the entity attributes of the head entity and the tail entity in the entity binary. The head entity in the entity binary is preferentially mapped to the entity space and projected to the relational space according to the entity space feature h of the tail entity in the entity binary and the mapping matrix Mr of the group relationship of the entity combination type corresponding to the entity binary, so as to obtain the relational space feature hr of the head entity in the entity binary. On the other hand, according to the entity space feature t of the tail entity in the entity binary and the mapping matrix Mr of the group relationship of the entity combination type corresponding to the entity binary, the head entity in the entity binary is preferentially mapped to the entity space and projected to the relational space to obtain the relational space feature tr of the tail entity in the entity binary. The relational space features of the head entity in the entity binary and the relational space features of the tail entity in the entity binary jointly form the relational space embedding features of the entity binary.
[0113] For example, the six entity tuples obtained by the combination processing in step 102 are: [policyholder A, insured B], [insured B, vehicle C], [policyholder A, vehicle C], [policyholder A, organization D], [insured B, organization D], and [vehicle C, organization D]. For the entity tuple [policyholder A, insured B], in step 1031, the entity representation layer in the trained knowledge graph embedding model is used to embed the entity attributes of the head entity (i.e., insured B) and the tail entity (i.e., insured B) in the entity tuple, and the entity attributes of the tail entity (i.e., insured B) are used to embed the entity representation layer in the trained knowledge graph. The head entity (i.e., policyholder A) and the tail entity (i.e., insured B) in the entity binary [policyholder A, insured B] are respectively represented by vectors to obtain the entity space features of the head entity (i.e., policyholder A) and the tail entity (i.e., insured B) in the entity binary [policyholder A, insured B]. Then, in step 1032, the trained knowledge graph is embedded in the relation conversion layer of the model. On the one hand, according to the head entity in the entity binary [policyholder A, insured B], the entity space features of the tail entity (i.e., insured B) are obtained. The entity space feature h of the entity (i.e., the insured A) and the mapping matrix Mr of the group relationship of the entity binary [insured A, insured B] corresponding to the entity combination type are used to map the head entity (i.e., the insured A) in the entity binary [insured A, insured B] to the entity space and project it to the relational space to obtain the relational space feature hr of the head entity (i.e., the insured A) in the entity binary [insured A, insured B]. On the other hand, according to the entity space feature h of the tail entity (i.e., the insured B) in the entity binary [insured A, insured B], the relationship space feature hr of the head entity (i.e., the insured A) is obtained. The mapping matrix Mr of the entity binary [policyholder A, insured B] corresponding to the entity combination type group relationship (i.e., the group relationship of combination type 1 [policyholder, insured]) is used to map the head entity (i.e., policyholder A) in the entity binary [policyholder A, insured B] to the entity space and project it to the relational space, and obtain the relational space feature tr of the tail entity (i.e., insured B) in the entity binary [policyholder A, insured B], thereby obtaining the relational space embedding feature of the entity binary [policyholder A, insured B]. By analogy, the relational space embedding features of the entity binary [policyholder A, insured B], [insured B, vehicle C], [policyholder A, vehicle C], [policyholder A, institution D], [insured B, institution D], and [vehicle C, institution D] are obtained respectively.
[0114] 104. Perform group relationship detection based on the relational space embedding features of the entity binary group using the trained link prediction model to obtain the group relationship detection result of the entity binary group.
[0115] In step 104, there are multiple ways to obtain the group detection result of the entity binary. For example, refer to Figure 3 , including the following methods ① and ②:
[0116] ① Using the relational space embedding features of entity binary groups for prediction. At this time, step 104 may specifically include the following steps 1041A to 1042A:
[0117] 1041A. Predictions are made based on the relational space embedding features of the entity binary using the trained link prediction model to obtain a group relationship score of the entity binary.
[0118] 1042A. Determine a gang detection result of the entity binary based on the gang relationship score.
[0119] Specifically, if the gang relationship score is greater than a preset score threshold, the existence of a gang relationship is used as the gang detection result of the entity binary; or, if the gang relationship score is less than or equal to the preset score threshold, the absence of a gang relationship is used as the gang detection result of the entity binary.
[0120] ② Use the relationship space embedding features and entity space embedding features of entity binary to make predictions. Figure 4 , step 104 may specifically include the following steps 1041B to 1043B:
[0121] 1041B. By embedding the trained knowledge graph into the entity representation layer of the model, vector representation is performed on each entity in the entity binary to obtain the entity space embedding feature of the entity binary.
[0122] The entity space embedding features include the entity space features of the head entity and the entity space features of the tail entity.
[0123] The implementation of step 1041B is similar to that of step 1031 , and details may be referred to the above related descriptions, which will not be repeated here.
[0124] 1042B. Predictions are made based on the relational space embedding features of the entity bigrams and the entity space embedding features of the entity bigrams using the trained link prediction model to obtain a group relationship score of the entity bigrams.
[0125] 1043B. Determine a gang detection result of the entity binary based on the gang relationship score.
[0126] The implementation of step 1043B is similar to that of step 1042A. For details, please refer to the relevant instructions in the previous text and will not be repeated here.
[0127] In steps 1041B to 1043B, on the one hand, by using the entity space embedding features of entity bigrams to predict gang relationship scores, the entity context semantics can be retained for prediction, thereby improving prediction accuracy; on the other hand, by combining the relational space embedding features of entity bigrams for prediction, the entity features can be mapped from the entity space to the relational space, thereby discovering potential gang relationships for collusion in fraud and improving prediction accuracy.
[0128] 105. Determine whether there is fraud in the target insurance case based on the gang relationship detection result of the entity binary group.
[0129] There are many ways to determine whether the target insurance case involves fraud in step 105, illustratively including:
[0130] (1) In step 105, when a gang relationship exists between entity binary tuples of one entity combination type, it is determined that fraud exists in the target insurance case. At this time, for the entity binary tuples of multiple entity combination types obtained by the combination processing in step 102, the gang detection results of the entity binary tuples of each entity combination type are obtained in step 104; in step 105, if the gang detection result of the entity binary tuple of one entity combination type is that a gang relationship exists, it is determined that fraud exists in the target insurance case; if the gang detection results of the entity binary tuples of multiple entity combination types are all that there is no gang relationship, it is determined that fraud does not exist in the target insurance case.
[0131] (2) In step 105, when entity binary tuples of multiple entity combination types have a gang relationship, it is determined that the target insurance case has fraud. At this time, for the entity binary tuples of multiple entity combination types obtained by the combination processing in step 102, the gang detection results of the entity binary tuples of each entity combination type are obtained in step 104; in step 105, if the gang detection result is that the number of entity binary tuples with a gang relationship is greater than the preset number threshold, it is determined that the target insurance case has fraud; if the gang detection result is that the number of entity binary tuples with a gang relationship is less than or equal to the preset number threshold, it is determined that the target insurance case has no fraud.
[0132] In order to better understand this embodiment, the following takes the insurance case of a car insurance case as an example to illustrate the insurance fraud detection process. Please refer to Figure 5 The fraud detection process of this insurance case includes:
[0133] 1. Build an insurance fraud knowledge graph.
[0134] For the specific implementation, please refer to the relevant instructions of step B2, which will not be repeated here.
[0135] 2. Use the constructed insurance fraud knowledge graph for training to obtain a trained knowledge graph embedding model. (Refer to Figure 5 "Insurance Fraud Knowledge Graph", "Knowledge Graph Embedding Model" section)
[0136] For the specific implementation, please refer to the relevant instructions of step B3, which will not be repeated here.
[0137] 3. Use the trained knowledge graph embedding model to construct new entity pair samples, train the preset link prediction model, and obtain the trained link prediction model. (Refer to Figure 5 "Relationship Inference", "Building Link Prediction Model", "Model Training and Evaluation" sections)
[0138] For the specific implementation, please refer to the relevant instructions of steps C1 to C3, which will not be repeated here.
[0139] 3. Obtain case data of the insurance cases to be tested (including vehicles, personnel, institutions, etc. involved).
[0140] 4. According to the preset entity combination type, entity combination processing is performed on each important entity involved in the insurance case to be tested to obtain the entity binary group of the insurance case to be tested.
[0141] For the specific implementation, please refer to the relevant description of step 102, which will not be repeated here.
[0142] 5. Use the trained knowledge graph embedding model to obtain the relational space embedding features of the entity binary of the insurance case to be tested.
[0143] For the specific implementation, please refer to the relevant description of step 103, which will not be repeated here.
[0144] 6. Using the trained link prediction model, the relationship space embedding features of the entity binary of the insurance case to be detected are used to predict and determine the gang relationship detection results of the entity binary of the insurance case to be detected. The blacklist of fraudulent behaviors can be dynamically updated based on the gang relationship detection results of the entity binary. (Refer to Figure 5 "Predicting Fraud Risk" and "Dynamically Updating Blacklists" sections)
[0145] For the specific implementation, please refer to the relevant description of step 104, which will not be repeated here.
[0146] 7. Determine whether there is fraud in the insurance case to be tested based on the gang relationship detection results of the entity binary group of the insurance case to be tested.
[0147] For the specific implementation, please refer to the relevant description of step 105, which will not be repeated here.
[0148] From the above content, it can be seen that, on the one hand, the relational space embedding features of the entity binary are obtained by mapping the entity binary to the relational space through the trained knowledge graph embedding model, and the gang relationship detection is obtained by using the relational space embedding features to obtain the gang relationship detection results of the entity binary. The entity can be mapped from the entity space to the relational space, thereby exploring the gang relationship of potential collusion fraud, avoiding the problem of inaccurate fraud detection due to the small amount of sample data of insurance cases with fraudulent behavior, and improving the detection accuracy of insurance fraud to a certain extent; on the other hand, based on the important entities involved in the target insurance case, entity binary groups are obtained through entity combination processing, and gang relationship detection can be performed on entity pairs of various combination types, and more entity pairs that may have fraud gang relationships in a single insurance case can be explored, so as to more comprehensively detect whether there is a gang relationship in a single insurance case and improve the detection accuracy of insurance fraud.
[0149] In addition, in order to better implement the insurance fraud behavior detection method in the embodiment of the present application, based on the insurance fraud behavior detection method, an insurance fraud behavior detection device is also provided in the embodiment of the present application, such as Figure 6 FIG. 1 is a schematic diagram of an embodiment of an insurance fraud detection device provided in an embodiment of the present application. The insurance fraud detection device 600 includes:
[0150] An acquisition unit 601 is used to acquire various important entities involved in a target insurance case and entity attributes of various important entities involved in the target insurance case;
[0151] A combination unit 602 is used to perform entity combination processing on each important entity involved in the target insurance case according to a preset entity combination type to obtain an entity binary group of the target insurance case;
[0152] The embedding unit 603 is used to obtain the relational space embedding features of the entity binary group according to the entity attributes of each important entity involved in the target insurance case through the trained knowledge graph embedding model;
[0153] A detection unit 604 is used to perform group relationship detection according to the relationship space embedding features of the entity binary group through the trained link prediction model to obtain the group relationship detection result of the entity binary group;
[0154] The detection unit 604 is further used to determine whether there is fraud in the target insurance case based on the gang relationship detection result of the entity binary group.
[0155] In some embodiments, the embedding unit 603 is specifically used for:
[0156] By embedding the trained knowledge graph into the entity representation layer of the model, according to the entity attributes of each important entity involved in the target insurance case, each entity in the entity binary is represented by a vector, and the entity space embedding feature of the entity binary is obtained;
[0157] By embedding the relationship conversion layer in the trained knowledge graph model, the relationship space conversion is performed according to the entity space embedding features of the entity binary and the group relationship conversion matrix corresponding to the relationship conversion layer to obtain the relationship space embedding features of the entity binary.
[0158] In some embodiments, the detection unit 604 is specifically used to:
[0159] Using the trained link prediction model, prediction is performed based on the relational space embedding features of the entity binary to obtain the group relationship score of the entity binary;
[0160] Based on the group relationship score, a group detection result of the entity binary is determined.
[0161] In some embodiments, the embedding unit 603 is specifically used for:
[0162] By using the trained entity representation layer in the knowledge graph embedding model, each entity in the entity binary is represented by a vector to obtain the entity space embedding feature of the entity binary, wherein the entity space embedding feature includes the entity space feature of the head entity and the entity space feature of the tail entity;
[0163] In some embodiments, the detection unit 604 is specifically used to:
[0164] The trained link prediction model is used to perform prediction based on the relational space embedding features of the entity bigrams and the entity space embedding features of the entity bigrams to obtain the group relationship score of the entity bigrams.
[0165] In some embodiments, the insurance fraud behavior detection device further includes a training unit (not shown in the figure), wherein the training unit is specifically used to:
[0166] Obtaining entity attributes of various important entities involved in the insurance case sample and entity relationships involved in the insurance case sample;
[0167] Constructing an insurance fraud knowledge graph based on entity attributes of various important entities involved in the insurance case samples and entity relationships involved in the insurance case samples;
[0168] According to the insurance fraud knowledge graph, a preset knowledge graph embedding model is trained to obtain a trained knowledge graph embedding model.
[0169] In some embodiments, the insurance fraud behavior detection device further includes a training unit (not shown in the figure), wherein the training unit is specifically used to:
[0170] Obtaining a first important entity from each important entity involved in the insurance case sample;
[0171] Using the trained knowledge graph embedding model, from each important entity involved in the insurance case sample, the second important entity with the highest probability of having a gang relationship with the first important entity is obtained;
[0172] The first important entity and the second important entity are taken as an entity pair having a gang relationship, and a preset link prediction model is trained to obtain a trained link prediction model.
[0173] In some embodiments, the insurance fraud behavior detection device further includes a screening unit (not shown in the figure), wherein the screening unit is specifically used to:
[0174] Acquire characteristic data of an insurance case sample, wherein the characteristic data includes a plurality of reference entities;
[0175] Training a preset fraud detection model according to the reference entity to obtain a trained fraud detection model;
[0176] Deleting the target reference entity in the feature data to obtain deleted feature data;
[0177] Using the deleted feature data to train a preset fraud detection model to obtain a retrained fraud detection model;
[0178] If the accuracy of the retrained fraud detection model is less than the accuracy of the trained fraud detection model, the target reference entity is taken as an important entity.
[0179] In specific implementation, the above units can be implemented as independent entities, or can be arbitrarily combined and implemented as the same or several entities. The specific implementation of the above units can refer to the above insurance fraud behavior detection method embodiment, which will not be repeated here.
[0180] See also Figure 7 , Figure 7 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application. The computer device may be a server.
[0181] like Figure 7 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.
[0182] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any insurance fraud detection method.
[0183] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.
[0184] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any insurance fraud detection method.
[0185] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0186] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.
[0187] In one embodiment, the processor is used to run a computer program stored in the memory to implement the following steps:
[0188] Obtain each important entity involved in a target insurance case, as well as the entity attributes of each important entity involved in the target insurance case; perform entity combination processing on each important entity involved in the target insurance case according to a preset entity combination type, to obtain an entity binary of the target insurance case; obtain the relational space embedding features of the entity binary according to the entity attributes of each important entity involved in the target insurance case through a trained knowledge graph embedding model; perform group relationship detection according to the relational space embedding features of the entity binary according to a trained link prediction model, to obtain the group relationship detection result of the entity binary; determine whether there is fraud in the target insurance case according to the group relationship detection result of the entity binary.
[0189] A person of ordinary skill in the art will appreciate that all or part of the steps in the above-mentioned insurance fraud detection method may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.
[0190] To this end, an embodiment of the present application provides a computer-readable storage medium, in which multiple computer programs are stored. The computer programs can be loaded by a processor to execute any insurance fraud detection method provided in the embodiment of the present application.
[0191] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.
[0192] In the above-mentioned insurance fraud detection device and computer-readable storage medium embodiments, the description of each embodiment has its own emphasis. For parts not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process and beneficial effects of the above-mentioned insurance fraud detection device, computer-readable storage medium and its corresponding units can refer to the description of the insurance fraud detection method in the above embodiment, and will not be repeated here.
[0193] The above is a detailed introduction to an insurance fraud detection method, device, computer equipment and computer-readable storage medium provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A method for detecting insurance fraud, characterized in that: The method comprises: Acquire various important entities involved in the target insurance case and entity attributes of various important entities involved in the target insurance case; According to a preset entity combination type, entity combination processing is performed on each important entity involved in the target insurance case to obtain an entity binary group of the target insurance case; By using the trained knowledge graph embedding model, the relational space embedding features of the entity binary are obtained according to the entity attributes of each important entity involved in the target insurance case; Using the trained link prediction model, group relationship detection is performed according to the relationship space embedding features of the entity binary group to obtain the group relationship detection result of the entity binary group; Based on the gang relationship detection result of the entity binary group, determine whether there is fraud in the target insurance case.
2. The insurance fraud detection method according to claim 1, characterized in that: The trained knowledge graph embedding model is used to obtain the relational space embedding features of the entity binary according to the entity attributes of each important entity involved in the target insurance case, including: By embedding the trained knowledge graph into the entity representation layer of the model, according to the entity attributes of each important entity involved in the target insurance case, each entity in the entity binary is represented by a vector, and the entity space embedding feature of the entity binary is obtained; By embedding the relationship conversion layer in the trained knowledge graph model, the relationship space conversion is performed according to the entity space embedding features of the entity binary and the group relationship conversion matrix corresponding to the relationship conversion layer to obtain the relationship space embedding features of the entity binary.
3. The insurance fraud detection method according to claim 1, characterized in that: The group relationship detection is performed according to the relationship space embedding features of the entity binary group using the trained link prediction model to obtain the group relationship detection result of the entity binary group, including: Using the trained link prediction model, prediction is performed based on the relational space embedding features of the entity binary to obtain the group relationship score of the entity binary; Based on the group relationship score, a group detection result of the entity binary is determined.
4. The insurance fraud detection method according to claim 3, characterized in that: The method further comprises: By using the trained entity representation layer in the knowledge graph embedding model, each entity in the entity binary is represented by a vector to obtain the entity space embedding feature of the entity binary, wherein the entity space embedding feature includes the entity space feature of the head entity and the entity space feature of the tail entity; The trained link prediction model is used to predict the entity binary group according to the relationship space embedding feature of the entity binary group to obtain the group relationship score of the entity binary group, including: The trained link prediction model is used to perform prediction based on the relational space embedding features of the entity bigrams and the entity space embedding features of the entity bigrams to obtain the group relationship score of the entity bigrams.
5. The insurance fraud detection method according to claim 1, characterized in that: The trained knowledge graph embedding model is obtained through the following steps: Obtaining entity attributes of various important entities involved in the insurance case sample and entity relationships involved in the insurance case sample; Constructing an insurance fraud knowledge graph based on entity attributes of various important entities involved in the insurance case samples and entity relationships involved in the insurance case samples; According to the insurance fraud knowledge graph, a preset knowledge graph embedding model is trained to obtain a trained knowledge graph embedding model.
6. The insurance fraud detection method according to claim 5, characterized in that: The method further comprises: Obtaining a first important entity from each important entity involved in the insurance case sample; Using the trained knowledge graph embedding model, from each important entity involved in the insurance case sample, the second important entity with the highest probability of having a gang relationship with the first important entity is obtained; The first important entity and the second important entity are taken as an entity pair having a gang relationship, and a preset link prediction model is trained to obtain a trained link prediction model.
7. The insurance fraud detection method according to claim 1, characterized in that: The method further comprises: Acquire characteristic data of an insurance case sample, wherein the characteristic data includes a plurality of reference entities; Training a preset fraud detection model according to the reference entity to obtain a trained fraud detection model; Deleting the target reference entity in the feature data to obtain deleted feature data; Using the deleted feature data to train a preset fraud detection model to obtain a retrained fraud detection model; If the accuracy of the retrained fraud detection model is less than the accuracy of the trained fraud detection model, the target reference entity is taken as an important entity.
8. An insurance fraud detection device, characterized in that: The insurance fraud behavior detection device comprises: An acquisition unit, used to acquire various important entities involved in a target insurance case and entity attributes of various important entities involved in the target insurance case; A combination unit, used to perform entity combination processing on each important entity involved in the target insurance case according to a preset entity combination type, to obtain an entity binary group of the target insurance case; An embedding unit, used to obtain the relational space embedding features of the entity binary group according to the entity attributes of each important entity involved in the target insurance case through a trained knowledge graph embedding model; A detection unit, configured to perform group relationship detection according to the relationship space embedding features of the entity binary group through a trained link prediction model, and obtain a group relationship detection result of the entity binary group; The detection unit is further used to determine whether there is fraud in the target insurance case based on the gang relationship detection result of the entity binary group.
9. A computer device, characterized in that: The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is used to execute the computer program and implement the insurance fraud detection method according to any one of claims 1 to 7 when executing the computer program.
10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is loaded by a processor to execute the insurance fraud detection method according to any one of claims 1 to 7.
