Detection model generation method and device, claim settlement analysis method and device, equipment and medium
By using policy network and value network training in the insurance claim detection model to generate the object detection model, combined with multiple request data from the same claim subject, the problem of insufficient utilization of manual annotation and correlation in the training of existing models is solved, and a higher level is achieved detection accuracy and intelligence.
Patent Information
- Application Number
- CN202411181007.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-20
AI Technical Summary
The existing insurance claims detection model requires manual labeling of data during training, and fails to effectively utilize the correlation between claims subjects, resulting in inaccurate and intelligent detection.
By obtaining claims request data and time tags, using policy network and value network models to train the object detection model, training the model based on unlabeled data, and training it in combination with multiple requests from the same claim subject to improve the model intelligence and detection accuracy .
It reduces the need for manpower labeling, lowers the threshold for model training, and improves the accuracy and intelligence of complex claims fraud detection.
Smart Images
Figure CN119168780A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of insurance claims processing, and in particular to a detection model generation method, claims analysis method, device, equipment and medium. Background Art
[0002] In the current field of insurance claims processing, reinforcement learning has been applied to address the possible fraud risks in group corporate claims, such as using the Deep-Q-Learning algorithm to train models to detect and analyze claims requests.
[0003] However, in the existing model training methods for group corporate claims, if the Deep-Q-Learning algorithm is used for model training, it is still necessary to manually label the results in advance (that is, mark whether the case is a fraud case). For example, when the model is learning a fraud scenario, if the data is not labeled and distinguished whether each data sample is a fraud case, then the model may be trained in the wrong direction.
[0004] On the other hand, since the existing detection models are trained in an isolated manner with respect to each data sample and do not consider the correlation between claims requests issued by the same claims entity, the generated detection models are not intelligent enough and are not accurate enough in detecting some complex claims frauds. Summary of the invention
[0005] The embodiments of the present application provide a detection model generation method, a claims analysis method, an apparatus, a device and a medium, which are intended to train a detection model that can be used to detect claims fraud based on unlabeled data, reduce the manpower required in the process of training the detection model, lower the model training threshold, and train the model in combination with multiple claims requests issued by the same claims subject in the data sample to improve the intelligence of the resulting model.
[0006] In a first aspect, an embodiment of the present application provides a detection model generation method, comprising:
[0007] Obtaining a data training set and retrieving a preset strategy network model and value network model, the data training set includes claim request data;
[0008] Parse the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests;
[0009] Based on the time labels, multiple claim requests of the same claim subject are input into the policy network model in sequence, and the claim behavior corresponding to each time label is determined according to the policy network model, where the claim subject is the individual who issues the claim request;
[0010] Input the claim settlement behavior into the value network model, and generate the behavior benefits corresponding to the claim settlement behavior according to the value network detection model;
[0011] Determine the fraud risk information corresponding to the claim request based on the claim behavior and behavior benefits;
[0012] Iteratively update the model parameters of the policy network model and the value network model according to the claim request and fraud risk information to generate a target policy network and a target value network;
[0013] A target detection model is generated based on the target strategy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
[0014] In some implementations, determining the claim settlement behavior corresponding to each time tag according to the policy network includes:
[0015] Obtain the individual financial status of the claim subject based on the time tag;
[0016] Based on the policy network model, the claim fraud probability is determined according to the individual financial status and the request characteristics of the claim request, and the claim behavior corresponding to the time label is determined according to the claim fraud probability.
[0017] In some implementations, the claim settlement behavior includes one of acceptance processing and case closing processing, and the claim settlement behavior corresponding to the time tag is determined according to the claim fraud probability, including:
[0018] Determine the expected value of claims of the claim subject according to the time tag and individual financial status, and calculate the comprehensive probability based on the expected value of claims and the probability of claims fraud;
[0019] If the comprehensive probability is greater than or equal to the preset probability threshold, the case is closed as the corresponding claim settlement action;
[0020] If the comprehensive probability is less than the preset probability threshold, the claim will be accepted and processed as the corresponding claim settlement behavior.
[0021] In some implementations, obtaining the individual financial status of the claim subject according to the time tag includes:
[0022] Obtain a preset number of historical time tags according to the time tags, and determine the historical claims behavior and historical behavior benefits corresponding to the historical time tags, where the time node corresponding to the historical time tag is before the time corresponding to the time tag;
[0023] Generate the cumulative income change characteristics of the claim subject based on historical behavioral income;
[0024] Determine the credit characteristics of the claim subject based on historical claim settlement behavior;
[0025] Determine individual financial status based on cumulative income change characteristics and credit characteristics.
[0026] In some embodiments, the model parameters of the policy network model include a probability distribution strategy, and the policy network model is used to determine the claim settlement behavior corresponding to each time tag according to the probability distribution strategy and the request characteristics of the claim settlement request;
[0027] Iterate and update the model parameters of the policy network model based on the claim request and fraud risk information, including:
[0028] Based on the policy network model, the request characteristics of the claim request corresponding to the current time tag are matched with the probability distribution strategy to obtain the target probability distribution;
[0029] Determine the claim settlement behavior and the behavior score of the claim settlement behavior based on the fraud risk information and the target probability distribution, and update the probability distribution strategy based on the behavior score;
[0030] The probability distribution strategy in the strategy network model is iteratively updated based on the claim requests corresponding to the subsequent time tags, and the scores of each behavior are counted to obtain the cumulative score until the cumulative score reaches the target score.
[0031] In some implementations, iteratively updating model parameters of the value network model according to the claim settlement request and the fraud risk information to generate a target value model includes:
[0032] Generate the expected benefit corresponding to the claim request according to the model parameters of the target value model, and assign the behavior benefit to the claim behavior corresponding to the claim request;
[0033] Calculate the difference between expected return and behavioral return;
[0034] The model parameters of the target value model are iteratively updated according to the profit difference until the profit difference is less than the target difference.
[0035] In a second aspect, the embodiment of the present application further provides a claims analysis method, the method comprising:
[0036] Calling a target detection model, where the target detection model is a target detection model generated by a detection model generation method provided in any embodiment of the specification of this application;
[0037] The request to be detected is input into the target detection model, so that the target detection model outputs corresponding fraud risk information as the claim analysis result according to the request to be detected.
[0038] In a third aspect, the present application also provides a detection model generation device, including:
[0039] A network model retrieval module is used to obtain a data training set and retrieve a preset strategy network model and value network model. The data training set includes claim request data.
[0040] A claim data parsing module is used to parse the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests;
[0041] The first model control module is used to input multiple claim requests of the same claim subject into the policy network model in sequence based on the time label, and determine the claim behavior corresponding to each time label according to the policy network model, wherein the claim subject is the individual who issues the claim request;
[0042] The second model control module is used to input the claim settlement behavior into the value network model and generate the behavior benefits corresponding to the claim settlement behavior according to the value network detection model;
[0043] A risk information generation module, used to determine the fraud risk information corresponding to the claim request according to the claim behavior and the behavior benefit;
[0044] A model parameter updating module, used for iteratively updating the model parameters of the policy network model and the value network model according to the claim request and fraud risk information to generate a target policy network and a target value network;
[0045] The target model generation module is used to generate a target detection model based on the target strategy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
[0046] In a fourth aspect, an embodiment of the present application further provides a computer device, the computer device comprising a memory and a processor;
[0047] Memory for storing computer programs;
[0048] A processor, used to execute a computer program and implement the detection model generation method provided in any embodiment of the present application specification and / or the claim risk analysis method provided in any embodiment of the present application specification when executing the computer program.
[0049] In a fifth aspect, an embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements a detection model generation method as provided in any embodiment of the present application specification and / or a claims risk analysis method as provided in any embodiment of the present application specification.
[0050] The embodiments of the present application provide a detection model generation method, a claims analysis method, an apparatus, a device and a medium. The detection model generation method includes: obtaining a data training set and calling a preset strategy network model and a value network model, wherein the data training set includes claims request data; parsing the claims request data to obtain multiple claims requests and time tags corresponding to the claims requests; inputting multiple claims requests of the same claims subject into the strategy network model in sequence based on the time tags, and determining the claims behavior corresponding to each time tag according to the strategy network model, wherein the claims subject is the individual who issues the claims request; inputting the claims behavior into the value network model, and generating the behavior benefit corresponding to the claims behavior according to the value network detection model; and Determine the fraud risk information corresponding to the claim request with the behavioral benefit; iteratively update the model parameters of the policy network model and the value network model according to the claim request and the fraud risk information to generate a target policy network and a target value network; generate a target detection model based on the target policy network and the target value network, the target detection model is used to obtain the request to be detected, and detect and generate the fraud risk information corresponding to the request to be detected, so as to train a detection model that can be used to detect claim fraud based on unlabeled data, reduce the manpower required in the process of training the detection model, lower the threshold for model training, and train the model in combination with multiple claim requests issued by the same claim subject in the data sample to improve the intelligence of the resulting model. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] 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.
[0052] Figure 1 A flowchart showing the steps of a detection model generation method provided in an embodiment of the present application;
[0053] Figure 2 A model framework diagram of a strategy network model and a value network model in a detection model generation method provided in an embodiment of the present application, and a model framework diagram of a target detection model;
[0054] Figure 3 A flowchart of the steps for determining a claim settlement behavior in a detection model generation method provided in an embodiment of the present application;
[0055] Figure 4 A flowchart of the steps of determining a claim settlement behavior according to a claim settlement fraud probability in a detection model generation method provided in an embodiment of the present application;
[0056] Figure 5A training flow chart of a strategy network model in a detection model generation method provided in an embodiment of the present application;
[0057] Figure 6 A training flow chart of a value network model in a detection model generation method provided in an embodiment of the present application;
[0058] Figure 7 A schematic diagram of the steps of a claims analysis method provided in an embodiment of the present application;
[0059] Figure 8 A schematic diagram of the module structure of a detection model generating device provided in an embodiment of the present application;
[0060] Fig. 9 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0061] 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 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 ordinary technicians in this field without creative work are within the scope of protection of this application.
[0062] 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.
[0063] In the current field of insurance claims processing, reinforcement learning has been applied to address the possible fraud risks in group corporate claims, such as using the Deep-Q-Learning algorithm to train models to detect and analyze claims requests.
[0064] However, in the existing model training methods for group corporate claims, if the Deep-Q-Learning algorithm is used for model training, it is still necessary to manually label the results in advance (that is, mark whether the case is a fraud case). For example, when the model is learning a fraud scenario, if the data is not labeled and distinguished whether each data sample is a fraud case, then the model may be trained in the wrong direction.
[0065] On the other hand, since the existing detection models are trained in an isolated manner with respect to each data sample and do not consider the correlation between claims requests issued by the same claims entity, the generated detection models are not intelligent enough and are not accurate enough in detecting some complex claims frauds.
[0066] Based on this, the embodiments of the present application provide a detection model generation method, a claims analysis method, an apparatus, a device and a medium, which aim to train a detection model that can be used to detect claims fraud based on unlabeled data, reduce the manpower required in the process of training the detection model, lower the model training threshold, and train the model in combination with multiple claims requests issued by the same claims entity in the data sample to improve the intelligence of the resulting model.
[0067] Among them, the detection model generation method can be applied to electronic devices such as computers, intelligent robots, independent servers or server clusters, and is not limited here.
[0068] In this embodiment, the detection model generation method is applied to an independent server as an example for description, but the detection model generation method is not limited to being used in only this scenario.
[0069] Some implementation modes 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 may be combined with each other.
[0070] See also Figure 1 , Figure 1 A flowchart illustrating the steps of a detection model generation method provided in an embodiment of the present application.
[0071] like Figure 1 As shown, the detection model generation method provided in the embodiment of the present application specifically includes the following steps S1-S7.
[0072] Step S1, obtaining a data training set and calling a preset strategy network model and a value network model, wherein the data training set includes claim request data.
[0073] Specifically, in an embodiment of the present application, a data training set is used as a training sample for a strategy network model and a value network model, wherein the data training set includes claim request data, and the claim request data includes multiple claim requests and time tags corresponding to the claim requests. The above multiple claim requests may be issued by the same claim subject or by at least two claim subjects.
[0074] It should be noted that the claim subject refers to the individual enterprise, enterprise group or individual that issues the claim request. It should also be noted that the time tag represents the time node when the claim request was issued.
[0075] See also Figure 2 , Figure 2 A model framework diagram of a strategy network model and a value network model in a detection model generation method provided in an embodiment of the present application, as well as a model framework diagram of a target detection model.
[0076] like Figure 2As shown in (a), the policy network model is used to obtain the input claim request and generate the claim behavior according to the claim request. The value network model is used to obtain the input claim behavior and generate the behavior benefit according to the claim request. Figure 2 (a) shows the model framework diagram of the policy network model and the value network model. After executing the steps of training the policy network model and the value network model in the detection model generation method, the target policy network and the target value network are obtained, for example Figure 2 The relationship between the target policy network and the target value network shown in (b).
[0077] Step S2: parsing the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests.
[0078] Specifically, after obtaining the data training set and calling the policy network model and the value network model, the above-mentioned multiple claim requests and the time tags corresponding to the claim requests are parsed from the claim request data, so as to train the policy network model and the value network model using the multiple claim requests and their corresponding time tags.
[0079] It should be noted that a claim request has the attribute of request characteristics. In some implementations, the policy network model determines the claim behavior corresponding to the claim request based on the request characteristics of the claim request.
[0080] Exemplarily, the request characteristics include but are not limited to at least one of claim type characteristics, claim amount characteristics, and claim cycle characteristics.
[0081] Step S3: Based on the time tags, multiple claim requests of the same claim subject are input into the policy network model in sequence, and the claim behavior corresponding to each time tag is determined according to the policy network model, wherein the claim subject is the individual who issues the claim request.
[0082] Specifically, first determine the claim entities corresponding to multiple claim requests, classify and filter out the claim requests issued by the same claim entity, and then input the claim requests issued by the same claim entity into the policy network model; specifically, input multiple claim requests of the same claim entity into the policy network model in sequence based on time labels, and determine the claim behavior corresponding to each time label according to the policy network model, so there is no need to manually label the data training set.
[0083] Specifically, the claim settlement behavior refers to the response behavior of the claim settlement object after a claim settlement request is issued to the claim settlement object at the time node corresponding to the time tag. The claim settlement object is, for example, a property insurance company.
[0084] It should be understood that by inputting multiple claim requests issued by the same claim subject into the policy network model in sequence based on time labels, the policy network model can generate claim behaviors in sequence according to the order of time labels, and in subsequent steps, train based on the time label order of multiple claim requests, thereby improving the intelligence of the resulting model, so as to accurately detect claim fraud for claim subjects that have issued multiple claim requests.
[0085] See also Figure 3 , Figure 3 A flowchart of the steps for determining claims settlement behavior in a detection model generation method provided in an embodiment of the present application.
[0086] like Figure 3 As shown, in some implementations, determining the claim settlement behavior corresponding to each time tag according to the policy network model in step S3 includes steps S31 to S32:
[0087] Step S31, obtaining the individual financial status of the claim subject according to the time tag;
[0088] Step S32: Based on the policy network model, the claim fraud probability is determined according to the individual financial status and the request characteristics of the claim request, and the claim behavior corresponding to the time tag is determined according to the claim fraud probability.
[0089] It should be noted that the above steps of determining the probability of claim fraud and determining the claim behavior corresponding to each time tag are based on the policy network model, and the server executing this method may input the time tag into the policy network model to perform the above operations.
[0090] Specifically, the policy network model first obtains the individual financial status of the claim subject based on the time label, then determines the claim fraud probability based on the individual financial status and the request characteristics of the claim request, and determines the claim behavior corresponding to the time label based on the claim fraud probability.
[0091] For example, individual financial status includes, but is not limited to, the asset structure, debt ratio, debt structure and profitability of the claims entity.
[0092] Specifically, the claim fraud probability is determined based on the individual financial status and the request characteristics of the claim request. The claim fraud probability represents the probability that the claim request is a claim fraud.
[0093] In some implementations, obtaining the individual financial status of the claim subject according to the time tag in step S31 includes:
[0094] Obtain a preset number of historical time tags according to the time tags, and determine the historical claims behavior and historical behavior benefits corresponding to the historical time tags, where the time node corresponding to the historical time tag is before the time corresponding to the time tag;
[0095] Generate the cumulative income change characteristics of the claim subject based on historical behavioral income;
[0096] Determine the credit characteristics of the claim subject based on historical claim settlement behavior;
[0097] Determine individual financial status based on cumulative income change characteristics and credit characteristics.
[0098] Among them, historical claims behavior refers to the response behavior of the claims object after a claim request is issued to the claims object at the time node corresponding to the historical time tag. Historical claims behavior includes acceptance processing and case closing processing. Acceptance processing refers to the acceptance of claim requests for payment, and case closing processing refers to the refusal to accept claim requests for payment.
[0099] It should be noted that historical behavior income refers to the net income obtained by the claim subject in the historical claims behavior corresponding to the historical time label. By statistically analyzing the historical behavior income corresponding to multiple historical time labels, the cumulative income of the claim subject under different historical time labels can be calculated to determine the cumulative income change characteristics of the claim subject.
[0100] In some embodiments, determining the credit characteristics of the claim subject based on historical claim settlement behaviors includes: determining the credit characteristics of the claim subject based on the number of historical claim settlement behaviors and the claim amount.
[0101] After determining the cumulative income change characteristics and credit characteristics, the actual individual financial status of the claimant can be accurately determined based on the cumulative income change characteristics and credit characteristics, so there is no need to manually mark the data samples.
[0102] In some embodiments, the claim handling behavior includes one of acceptance processing and case closing processing. Acceptance processing refers to accepting the claim request for payment, and case closing processing refers to refusing to accept the claim request for payment.
[0103] See also Figure 4 , Figure 4 A flow chart of the steps of determining claim behavior based on the probability of claim fraud in a detection model generation method provided in an embodiment of the present application.
[0104] like Figure 4 As shown, correspondingly, in step S32, the claim settlement behavior corresponding to the time tag is determined according to the claim settlement fraud probability, including steps S321 to S323:
[0105] Step S321: determining the expected value of the claim subject according to the time tag and the individual financial status, and calculating and obtaining the comprehensive probability according to the expected value of the claim and the probability of claim fraud;
[0106] Step S322: if the comprehensive probability is greater than or equal to the preset probability threshold, then the case closing process is determined as the corresponding claim settlement action;
[0107] Step S323: If the comprehensive probability is less than the preset probability threshold, the acceptance process is determined as the corresponding claim settlement behavior.
[0108] Specifically, the expected value of the claim subject refers to the claim amount determined by the strategy network model based on the current model parameters, time labels and the individual financial status of the claim subject. The comprehensive probability refers to the probability obtained by correcting the probability of claim fraud according to the expected value of the claim, which represents the probability that the claim request is a claim fraud.
[0109] After obtaining the comprehensive probability, if the comprehensive probability is greater than or equal to the preset probability threshold, the case closure process is determined as the corresponding claim settlement behavior; if the comprehensive probability is less than the preset probability threshold, the case acceptance process is determined as the corresponding claim settlement behavior.
[0110] Step S4: input the claim settlement behavior into the value network model, and generate the behavior benefit corresponding to the claim settlement behavior according to the value network detection model.
[0111] It should be noted that the above step of generating the behavior benefits corresponding to the claim settlement behavior is based on the value network model, and the server executing this method may input the claim settlement behavior into the strategy network model to perform the above operation. Specifically, the value network model is used to receive the input claim settlement behavior, and generate the behavior benefits corresponding to the claim settlement behavior based on the claim settlement behavior and the current model parameters of the value network model.
[0112] Step S5: Determine the fraud risk information corresponding to the claim settlement request based on the claim settlement behavior and the behavior benefit.
[0113] Specifically, after executing the above steps, for each time label, the strategy network model generates the corresponding claim behavior, and the value network model generates the corresponding behavior benefit. After that, the fraud risk information corresponding to a single time label can be obtained based on the claim behavior and behavior benefit corresponding to the same time label.
[0114] It should be noted that, from the perspective of behavioral benefits, the higher the behavioral benefits, the higher the fraud probability represented by the fraud risk information. From the perspective of claims settlement behavior, the more times the claims settlement behavior is accepted and processed within a preset period of time, the higher the fraud probability represented by the fraud risk information.
[0115] Step S6: iteratively update the model parameters of the policy network model and the value network model according to the claim settlement request and the fraud risk information to generate a target policy network and a target value network.
[0116] See also Figure 5 , Figure 5 A training flowchart of a strategy network model in a detection model generation method provided in an embodiment of the present application.
[0117] like Figure 5 As shown, in some embodiments, the model parameters of the policy network model include a probability distribution strategy, and the policy network model is used to determine the claim behavior corresponding to each time tag according to the probability distribution strategy and the request characteristics of the claim request;
[0118] Iterate and update the model parameters of the policy network model based on the claim request and fraud risk information, including:
[0119] Based on the policy network model, the request characteristics of the claim request corresponding to the current time tag are matched with the probability distribution strategy to obtain the target probability distribution;
[0120] Determine the claim settlement behavior and the behavior score of the claim settlement behavior based on the fraud risk information and the target probability distribution, and update the probability distribution strategy based on the behavior score;
[0121] The probability distribution strategy in the strategy network model is iteratively updated based on the claim requests corresponding to the subsequent time tags, and the scores of each behavior are counted to obtain the cumulative score until the cumulative score reaches the target score.
[0122] It should be noted that the probability distribution strategy refers to: selecting the corresponding target probability distribution for various claim requests with different request characteristics.
[0123] After matching the target probability distribution corresponding to the request feature, the claim behavior and the behavior score of the claim behavior are determined according to the fraud risk information and the target probability distribution, and the probability distribution strategy is updated according to the behavior score; the probability distribution strategy in the policy network model is iteratively updated based on the claim request corresponding to the subsequent time tag, and the behavior score is counted to obtain the cumulative score until the cumulative score reaches the target score. It is determined that the training of the policy network model is completed, and the current policy network model parameters have met the conditions for completing the training, and the trained policy network model is used as the target policy network.
[0124] like Figure 5As shown, it should be noted that iteratively updating the probability distribution strategy in the policy network model based on the claim requests corresponding to subsequent time tags means: based on the time sequence of the corresponding time tags, a number of claim requests are successively input into the policy network model to be trained. For example, the claim request with the earliest time tag is first input into the policy network model A1, and the policy network model A1 determines the target probability distribution based on the current probability distribution strategy a1, and generates a claim behavior and a corresponding behavior score 1. Then, according to the behavior score 1, the probability distribution strategy a1 is updated to the probability distribution strategy a2, so that the policy network model A1 learns to be the policy network model A2, and the claim request with the second highest order in terms of time tags is input into the policy network model A2. The policy network model A2 determines the target probability distribution based on the current probability distribution strategy a2, and generates a claim behavior and a corresponding behavior score 2.
[0125] Similarly, the scores of each behavior are accumulated to obtain the cumulative score. If the cumulative score reaches the target score when the policy network model is iteratively updated to the policy network model An, it is determined that the policy network model An has reached the conditions for completing the training, and the iterative update of the policy network model is completed, and the policy network model An is used as the target policy network.
[0126] It should be noted that the behavior score represents the anti-fraud effect of the claim behavior on the claim request. The detection model training method provided in the embodiment of the present application aims to train a detection model that can accurately detect claim fraud for a single claim entity. Therefore, during the training process, when the cumulative score reaches the target score, it is confirmed that the model parameters of the current policy network model have met the conditions for completing the training, and the trained policy network model is used as the target policy network.
[0127] See also Figure 6 , Figure 6 A training flowchart of a value network model in a detection model generation method provided in an embodiment of the present application.
[0128] like Figure 6 As shown, in some embodiments, the model parameters of the value network model are iteratively updated according to the claim settlement request and the fraud risk information to generate a target value model, including:
[0129] Generate the expected benefit corresponding to the claim request according to the model parameters of the target value model, and assign the behavior benefit to the claim behavior corresponding to the claim request;
[0130] Calculate the difference between expected return and behavioral return;
[0131] The model parameters of the target value model are iteratively updated according to the profit difference until the profit difference is less than the target difference.
[0132] Specifically, the server executing this method calls the value network model, generates the expected profit corresponding to the claim request based on the model parameters of the target value model, and assigns the behavioral profit to the claim behavior corresponding to the claim request. The server executing this method calculates the profit difference between the expected profit and the behavioral profit, and iteratively updates the model parameters of the target value model according to the profit difference until the profit difference is less than the target difference.
[0133] It should be noted that the iterative update of the policy network model and the value network model can be performed simultaneously, or the model parameters of the policy network model can be iteratively updated first, and then the model parameters of the value network model can be iteratively updated after the training of the policy network model is completed. If the iterative update of the policy network model and the value network model is performed simultaneously, after the training of one of the policy network model and the value network model is completed, the training of the other one will continue until the conditions for completing the training are met.
[0134] Step S7: Generate a target detection model based on the target strategy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
[0135] Specifically, the server executing this method packages the target policy network and the target value network into models to generate the required target detection model.
[0136] like Figure 2 As shown in (b), the target detection model includes a target strategy network and a target value network. The target strategy network is used to obtain the request to be detected and generate a claim behavior based on the request to be detected. The target value network is used to generate corresponding behavior benefits based on the claim behavior. The target detection model is used to output the above-mentioned claim behavior and behavior benefits as fraud risk information.
[0137] See also Figure 7 , Figure 7 A schematic diagram of the steps of a claims analysis method provided in an embodiment of the present application.
[0138] like Figure 2 (b) with Figure 7 As shown, the claim analysis method provided in the embodiment of the present application specifically includes steps S8 to S9:
[0139] Step S8: calling a target detection model, where the target detection model is a target detection model generated by a detection model generation method provided in any embodiment of the specification of this application;
[0140] Step S9: Input the request to be detected into the target detection model, so that the target detection model outputs corresponding fraud risk information as the claim analysis result according to the request to be detected.
[0141] Specifically, the server executing this method first calls a target detection model, which is a target detection model generated by a detection model generation method such as that provided in any embodiment of the present application specification, and then inputs the request to be detected into the target detection model, so that the target detection model outputs corresponding fraud risk information based on the request to be detected, and uses the fraud risk information as the claim analysis result.
[0142] It should be noted that the target detection model is used to output fraud risk information based on the request to be detected, and the fraud risk information, that is, the claim analysis result includes the claim behavior and behavior benefits corresponding to the request to be detected, and the request to be detected can be processed based on the claim behavior and behavior benefits.
[0143] See also Figure 8 , Figure 8 A schematic diagram of the module structure of a detection model generating device provided in an embodiment of the present application.
[0144] like Figure 8 As shown, the detection model generating device 100 includes:
[0145] The network model retrieval module 101 is used to obtain a data training set and retrieve a preset strategy network model and a value network model, wherein the data training set includes claim request data;
[0146] The claim data parsing module 102 is used to parse the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests;
[0147] The first model control module 103 is used to input multiple claim requests of the same claim subject into the policy network model in sequence based on the time label, and determine the claim behavior corresponding to each time label according to the policy network model, wherein the claim subject is the individual who issues the claim request;
[0148] The second model control module 104 is used to input the claim settlement behavior into the value network model and generate the behavior benefit corresponding to the claim settlement behavior according to the value network detection model;
[0149] The risk information generating module 105 is used to determine the fraud risk information corresponding to the claim settlement request according to the claim settlement behavior and the behavior benefit;
[0150] A model parameter updating module 106, for iteratively updating the model parameters of the policy network model and the value network model according to the claim settlement request and the fraud risk information to generate a target policy network and a target value network;
[0151] The target model generation module 107 is used to generate a target detection model based on the target policy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
[0152] In some implementations, the first model control module 103 determines the claim settlement behavior corresponding to each time tag according to the policy network, including:
[0153] Obtain the individual financial status of the claim subject based on the time tag;
[0154] Based on the policy network model, the claim fraud probability is determined according to the individual financial status and the request characteristics of the claim request, and the claim behavior corresponding to the time label is determined according to the claim fraud probability.
[0155] In some embodiments, the claim handling behavior includes one of acceptance processing and case closing processing;
[0156] The first model control module 103 determines the claim settlement behavior corresponding to the time tag according to the claim settlement fraud probability, including:
[0157] Determine the expected value of claims of the claim subject according to the time tag and individual financial status, and calculate the comprehensive probability based on the expected value of claims and the probability of claims fraud;
[0158] If the comprehensive probability is greater than or equal to the preset probability threshold, the case is closed as the corresponding claim settlement action;
[0159] If the comprehensive probability is less than the preset probability threshold, the claim will be accepted and processed as the corresponding claim settlement behavior.
[0160] In some implementations, the first model control module 103 obtains the individual financial status of the claim subject according to the time tag, including:
[0161] Obtain a preset number of historical time tags according to the time tags, and determine the historical claims behavior and historical behavior benefits corresponding to the historical time tags, where the time node corresponding to the historical time tag is before the time corresponding to the time tag;
[0162] Generate the cumulative income change characteristics of the claim subject based on historical behavioral income;
[0163] Determine the credit characteristics of the claim subject based on historical claim settlement behavior;
[0164] Determine individual financial status based on cumulative income change characteristics and credit characteristics.
[0165] In some embodiments, the model parameters of the policy network model include a probability distribution strategy, and the policy network model is used to determine the claim settlement behavior corresponding to each time tag according to the probability distribution strategy and the request characteristics of the claim settlement request;
[0166] The model parameter updating module 106 iteratively updates the model parameters of the policy network model according to the claim settlement request and the fraud risk information, including:
[0167] Based on the policy network model, the request characteristics of the claim request corresponding to the current time tag are matched with the probability distribution strategy to obtain the target probability distribution;
[0168] Determine the claim settlement behavior and the behavior score of the claim settlement behavior based on the fraud risk information and the target probability distribution, and update the probability distribution strategy based on the behavior score;
[0169] The probability distribution strategy in the strategy network model is iteratively updated based on the claim requests corresponding to the subsequent time tags, and the scores of each behavior are counted to obtain the cumulative score until the cumulative score reaches the target score.
[0170] In some implementations, the model parameter updating module 106 iteratively updates the model parameters of the value network model according to the claim settlement request and the fraud risk information to generate a target value model, including:
[0171] Generate the expected benefit corresponding to the claim request according to the model parameters of the target value model, and assign the behavior benefit to the claim behavior corresponding to the claim request;
[0172] Calculate the difference between expected return and behavioral return;
[0173] The model parameters of the target value model are iteratively updated according to the profit difference until the profit difference is less than the target difference.
[0174] See also Fig. 9 , Fig. 9 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application.
[0175] like Fig. 9 As shown, the computer device 200 includes a processor 201 and a memory 202, and the processor 201 and the memory 202 are connected via a bus 203, such as an I2C (Inter-Integrated Circuit) bus.
[0176] Specifically, the processor 201 is used to provide computing and control capabilities to support the operation of the entire computer device. The processor 201 can be a central processing unit (CPU), and the processor 201 can 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 gates or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.
[0177] Specifically, the memory 702 may be a Flash chip, a read-only memory (ROM) disk, an optical disk, a USB disk, or a mobile hard disk.
[0178] Those skilled in the art will understand that Figure 7 The structure shown in the figure is merely a block diagram of a partial structure related to the embodiment of the present application, and does not constitute a limitation on the computer device to which the embodiment 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.
[0179] Among them, the processor is used to run a computer program stored in the memory, and implement any one of the detection model generation methods provided in the embodiments of the present application and / or any one of the claims risk analysis methods provided in the embodiments of the present application specification when executing the computer program.
[0180] In some implementations, the processor 201 is configured to run a computer program stored in the memory 202, and implement the following steps when executing the computer program:
[0181] Obtaining a data training set and retrieving a preset strategy network model and value network model, the data training set includes claim request data;
[0182] Parse the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests;
[0183] Based on the time labels, multiple claim requests of the same claim subject are input into the policy network model in sequence, and the claim behavior corresponding to each time label is determined according to the policy network model, where the claim subject is the individual who issues the claim request;
[0184] Input the claim settlement behavior into the value network model, and generate the behavior benefits corresponding to the claim settlement behavior according to the value network detection model;
[0185] Determine the fraud risk information corresponding to the claim request based on the claim behavior and behavior benefits;
[0186] Iteratively update the model parameters of the policy network model and the value network model according to the claim request and fraud risk information to generate a target policy network and a target value network;
[0187] A target detection model is generated based on the target strategy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
[0188] In some implementations, the processor 201 determines the claim settlement behavior corresponding to each time tag according to the policy network, including:
[0189] Obtain the individual financial status of the claim subject based on the time tag;
[0190] Based on the policy network model, the claim fraud probability is determined according to the individual financial status and the request characteristics of the claim request, and the claim behavior corresponding to the time label is determined according to the claim fraud probability.
[0191] In some embodiments, the claim handling behavior includes one of acceptance processing and case closing processing;
[0192] The processor 201 determines the claim settlement behavior corresponding to the time tag according to the claim settlement fraud probability, including:
[0193] Determine the expected value of claims of the claim subject according to the time tag and individual financial status, and calculate the comprehensive probability based on the expected value of claims and the probability of claims fraud;
[0194] If the comprehensive probability is greater than or equal to the preset probability threshold, the case is closed as the corresponding claim settlement action;
[0195] If the comprehensive probability is less than the preset probability threshold, the claim will be accepted and processed as the corresponding claim settlement behavior.
[0196] In some implementations, the processor 201 obtains the individual financial status of the claim subject according to the time tag, including:
[0197] Obtain a preset number of historical time tags according to the time tags, and determine the historical claims behavior and historical behavior benefits corresponding to the historical time tags, where the time node corresponding to the historical time tag is before the time corresponding to the time tag;
[0198] Generate the cumulative income change characteristics of the claim subject based on historical behavioral income;
[0199] Determine the credit characteristics of the claim subject based on historical claim settlement behavior;
[0200] Determine individual financial status based on cumulative income change characteristics and credit characteristics.
[0201] In some embodiments, the model parameters of the policy network model include a probability distribution strategy, and the policy network model is used to determine the claim settlement behavior corresponding to each time tag according to the probability distribution strategy and the request characteristics of the claim settlement request;
[0202] The processor 201 iteratively updates the model parameters of the policy network model according to the claim settlement request and the fraud risk information, including:
[0203] Based on the policy network model, the request characteristics of the claim request corresponding to the current time tag are matched with the probability distribution strategy to obtain the target probability distribution;
[0204] Determine the claim settlement behavior and the behavior score of the claim settlement behavior based on the fraud risk information and the target probability distribution, and update the probability distribution strategy based on the behavior score;
[0205] The probability distribution strategy in the strategy network model is iteratively updated based on the claim requests corresponding to the subsequent time tags, and the scores of each behavior are counted to obtain the cumulative score until the cumulative score reaches the target score.
[0206] In some implementations, the processor 201 iteratively updates the model parameters of the value network model according to the claim settlement request and the fraud risk information to generate a target value model, including:
[0207] Generate the expected benefit corresponding to the claim request according to the model parameters of the target value model, and assign the behavior benefit to the claim behavior corresponding to the claim request;
[0208] Calculate the difference between expected return and behavioral return;
[0209] The model parameters of the target value model are iteratively updated according to the profit difference until the profit difference is less than the target difference.
[0210] In some implementations, the processor 201 is configured to run a computer program stored in the memory 202, and implement the following steps when executing the computer program:
[0211] Calling a target detection model, where the target detection model is a target detection model generated by a detection model generation method provided in any embodiment of the specification of this application;
[0212] The request to be detected is input into the target detection model, so that the target detection model outputs corresponding fraud risk information as the claim analysis result according to the request to be detected.
[0213] It should be noted that those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, the specific working process of the computer device described above can refer to the corresponding process in the aforementioned detection model generation method embodiment, and will not be repeated here.
[0214] An embodiment of the present application also provides a storage medium, which stores a computer program. The computer program can be executed by one or more processors to implement any detection model generation method provided in the embodiment of the present application and / or any claim risk analysis method provided in the embodiment of the present application specification.
[0215] The storage medium may be an internal storage unit of the computer device described in the foregoing embodiment, such as a hard disk or memory of the computer device. The storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Med ia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc., equipped on the computer device.
[0216] In summary, the embodiments of the present application provide a detection model generation method, a claims analysis method, an apparatus, a device and a medium. The detection model generation method includes: obtaining a data training set and calling a preset strategy network model and a value network model, wherein the data training set includes claims request data; parsing the claims request data to obtain multiple claims requests and time tags corresponding to the claims requests; inputting multiple claims requests of the same claims subject into the strategy network model in sequence based on the time tags, and determining the claims behavior corresponding to each time tag according to the strategy network model, wherein the claims subject is the individual who issues the claims request; inputting the claims behavior into the value network model, and generating the behavior benefit corresponding to the claims behavior according to the value network detection model; and Determine the fraud risk information corresponding to the claim request and the behavioral benefit; iteratively update the model parameters of the policy network model and the value network model according to the claim request and the fraud risk information to generate a target policy network and a target value network; generate a target detection model based on the target policy network and the target value network, and the target detection model is used to obtain the request to be detected, and detect and generate the fraud risk information corresponding to the request to be detected, so as to train a detection model that can be used to detect claim fraud based on unlabeled data, reduce the manpower required in the process of training the detection model, lower the threshold for model training, and train the model in combination with multiple claim requests issued by the same claim subject in the data sample to improve the intelligence of the resulting model.
[0217] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In a hardware embodiment, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some 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 implemented as hardware, or 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 non-transitory medium) and a communication medium (or 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 technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that communication media typically contain 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.
[0218] In the description of this application, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.
[0219] It should be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0220] It should also be understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations. It should be noted that, in this article, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "including a..." does not exclude the presence of other identical elements in the process, method, article or system including the element.
[0221] The serial numbers of the embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments. The above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.
Claims
1. A detection model generation method, characterized in that: The method comprises: Obtaining a data training set and retrieving a preset strategy network model and a value network model, wherein the data training set includes claim request data; Parsing the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests; Based on the time tags, the multiple claim requests of the same claim subject are sequentially input into the policy network model, and the claim actions corresponding to the time tags are determined according to the policy network model, wherein the claim subject is the individual who issues the claim request; Inputting the claim settlement behavior into the value network model, and generating the behavior benefit corresponding to the claim settlement behavior according to the value network detection model; Determining fraud risk information corresponding to the claim settlement request based on the claim settlement behavior and the behavior benefits; Iteratively updating the model parameters of the policy network model and the value network model according to the claim settlement request and the fraud risk information to generate a target policy network and a target value network; A target detection model is generated based on the target strategy network and the target value network. The target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
2. The detection model generation method according to claim 1, characterized in that: The determining, according to the policy network, the claim settlement behavior corresponding to each of the time tags includes: Obtaining the individual financial status of the claim subject according to the time tag; Based on the policy network model, the claim fraud probability is determined according to the individual financial status and the request characteristics of the claim request, and the claim behavior corresponding to the time label is determined according to the claim fraud probability.
3. The detection model generation method according to claim 2, characterized in that: The claim settlement behavior includes one of acceptance processing and case closing processing, and the claim settlement behavior corresponding to the time tag is determined according to the claim fraud probability, including: Determine the expected value of the claim settlement of the claim settlement subject according to the time tag and the individual financial status, and calculate and obtain a comprehensive probability according to the expected value of the claim settlement and the probability of claim settlement fraud; If the comprehensive probability is greater than or equal to the preset probability threshold, then the case is closed as the corresponding claim settlement action; If the comprehensive probability is less than the preset probability threshold, the acceptance process is determined as the corresponding claim settlement behavior.
4. The detection model generation method according to claim 2, characterized in that: The obtaining of the individual financial status of the claim subject according to the time tag includes: Obtain a preset number of historical time tags according to the time tags, and determine the historical claims settlement behaviors and historical behavior benefits corresponding to the historical time tags, where the time nodes corresponding to the historical time tags are before the time corresponding to the time tags; Generate cumulative income change characteristics of the claim subject based on the historical behavior income; Determining the credit characteristics of the claim settlement entity based on the historical claim settlement behavior; The individual financial status is determined according to the cumulative income change characteristics and the credit characteristics.
5. The detection model generation method according to any one of claims 1 to 4, characterized in that: The model parameters of the strategy network model include a probability distribution strategy, and the strategy network model is used to determine the claim settlement behavior corresponding to each of the time tags according to the probability distribution strategy and the request characteristics of the claim settlement request; The iterative updating of the model parameters of the policy network model according to the claim settlement request and the fraud risk information includes: Based on the policy network model, matching the request feature of the claim settlement request corresponding to the current time tag with the probability distribution strategy to obtain a target probability distribution; Determining the claim settlement behavior and the behavior score of the claim settlement behavior according to the fraud risk information and the target probability distribution, and updating the probability distribution strategy according to the behavior score; The probability distribution strategy in the strategy network model is iteratively updated based on the subsequent claim settlement request corresponding to the time tag, and the scores of each behavior are counted to obtain a cumulative score until the cumulative score reaches the target score.
6. The detection model generation method according to claim 5, characterized in that: The iterative updating of the model parameters of the value network model according to the claim settlement request and the fraud risk information to generate a target value model includes: generating an expected benefit corresponding to the claim settlement request according to the model parameters of the target value model, and assigning the behavior benefit to the claim settlement behavior corresponding to the claim settlement request; Calculating the difference between the expected benefit and the behavioral benefit; The model parameters of the target value model are iteratively updated according to the profit difference until the profit difference is less than the target difference.
7. A claims analysis method, characterized in that: The method comprises: Calling a target detection model, wherein the target detection model is a target detection model generated by the detection model generation method according to any one of claims 1 to 6; The request to be detected is input into the target detection model, so that the target detection model outputs corresponding fraud risk information as a claim analysis result according to the request to be detected.
8. A detection model generation device, characterized in that: include: A network model retrieval module, used to obtain a data training set and retrieve a preset strategy network model and a value network model, wherein the data training set includes claim request data; A claim data parsing module, used to parse the claim request data to obtain multiple claim requests and time tags corresponding to the claim requests; A first model control module, which sequentially inputs a plurality of the claim settlement requests of the same claim settlement subject into the policy network model based on the time label, and determines the claim settlement behavior corresponding to each of the time labels according to the policy network model, wherein the claim settlement subject is the individual who issues the claim settlement request; A second model control module is used to input the claim settlement behavior into the value network model, and generate a behavior benefit corresponding to the claim settlement behavior according to the value network detection model; A risk information generating module, used to determine the fraud risk information corresponding to the claim settlement request according to the claim settlement behavior and the behavior benefit; A model parameter updating module, used for iteratively updating the model parameters of the policy network model and the value network model according to the claim settlement request and the fraud risk information to generate a target policy network and a target value network; The target model generation module is used to generate a target detection model based on the target strategy network and the target value network, and the target detection model is used to obtain the request to be detected and detect and generate fraud risk information corresponding to the request to be detected.
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 detection model generation method as described in any one of claims 1 to 6 and / or the claim risk analysis method as described in claim 7 when executing the computer program.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to implement the detection model generation method according to any one of claims 1 to 6 and / or the claim risk analysis method according to claim 7.