Insurance claim processing method, device, electronic device and storage medium

Abstract

The present application provides a method, device, electronic device and storage medium for processing insurance policy claims, which are applied in the computer field. The method includes: obtaining the insurance policy information to be processed; inputting the insurance policy information into the claims strategy perception machine for processing to obtain the first claims identification module score, and inputting the insurance policy information into the multi-layer perceptron for processing to obtain the second claims identification module score, wherein the claims strategy perception machine is trained based on the loss function of the imbalanced loss function; and performing claims operations on the insurance policy information based on the claims review strategy that matches the first claims identification module score and the second claims identification module score. This solution reduces the huge demand for human resources in the claims process, and improves the efficiency of claims while ensuring the accuracy of claim recognition.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a method, device, electronic device and storage medium for processing insurance policy claims. Background Art

[0002] At present, most of the existing claims are based on manual claims investigation by business personnel, which is inefficient. Moreover, this method relies too much on manual investigation by business personnel, which will cause a huge human resource gap. The different working abilities of different business personnel, the inconsistent review levels, and the overly subjective review can easily bring unnecessary losses to the claims process. Second, if the rejected claims cannot be effectively identified, it will also bring additional compensation risks to the company. Some related models currently used in the industry mostly use default parameters and are not very close to the business. Summary of the invention

[0003] The present application provides a method for processing insurance policy claims, the method comprising:

[0004] Get pending policy information;

[0005] Input the policy information into a claims strategy perceptron for processing to obtain a first claims identification module score, and input the policy information into a multilayer perceptron for processing to obtain a second claims identification module score, wherein the claims strategy perceptron is trained based on a loss function of an imbalanced loss function;

[0006] A claims verification operation is performed on the insurance policy information based on a claims verification review strategy in which the first claims verification identification module score matches the second claims verification identification module score.

[0007] Optionally, the claims settlement strategy perceptron is trained through the following steps:

[0008] Obtain sample policy information and build an initialized Light GBM model;

[0009] Get the first and second derivatives of the binary cross entropy function in the Light GBM model;

[0010] The gradient modulus of the sample policy information is calculated based on the first-order derivative and the second-order derivative;

[0011] Transforming the gradient modulus according to a preset density parameter to obtain a target gradient density;

[0012] Constructing a target loss function based on an imbalanced loss function according to the target gradient density;

[0013] The sample policy information is used to train the Light GBM model deployed with the target loss function to obtain a claims settlement strategy perceptron.

[0014] Optionally, after constructing a target loss function based on an imbalanced loss function according to the target gradient density, the method includes:

[0015] The target loss function is adjusted by the accumulated value of the sample quantity of the sample policy information.

[0016] Optionally, transforming the gradient modulus according to a preset density parameter to obtain a target gradient density includes:

[0017] ;

[0018] ;

[0019] Wherein, the GD(g) is the target gradient density, the ( ) is the number of samples in a unit area, is the unit area length, which can be as small as close to 0, g is the gradient modulus of the sample, and x is the number of samples in the unit area.

[0020] Optionally, the adjusting the target loss function by the accumulated value of the sample quantity of the sample policy information includes:

[0021] ;

[0022] ;

[0023] Among them, the is the inverse of the gradient density of the i-th sample, N is the total number of samples, is the gradient modulus of the i-th sample, is the unit area length, which can be as small as close to 0, GD(g) is the target gradient density, ( ) is the number of samples in a unit area, is the adjusted objective loss function, the CE ( , ) is the binary cross entropy loss function of the i-th sample, is the sample category probability output by the activation function of the i-th sample, is the category value of the i-th sample.

[0024] Optionally, the claim verification strategy based on the matching of the first claim verification identification module score and the second claim verification identification module score performs a claim verification operation on the policy information, including:

[0025] Combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score;

[0026] When the comprehensive claim identification score is less than the claim identification score threshold, the policy information is submitted to manual claim verification;

[0027] When the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the policy information is submitted to the intelligent claim model for claim settlement.

[0028] Optionally, obtaining sample policy information includes:

[0029] Obtain initial policy information;

[0030] The initial insurance policy is processed according to a target preprocessing method to obtain sample insurance policy information, and the target preprocessing method includes: at least one of: outlier processing, data visualization analysis, and data standardization.

[0031] The present application provides a device for processing insurance claims, the device comprising:

[0032] A receiving module, configured to obtain the insurance policy information to be processed;

[0033] The prediction module is configured to input the policy information into a claims strategy perceptron for processing to obtain a first claims identification module score, and input the policy information into a multilayer perceptron for processing to obtain a second claims identification module score, wherein the claims strategy perceptron is trained based on a loss function of an imbalanced loss function;

[0034] The processing module is configured to perform a claim verification operation on the insurance policy information based on a claim verification review strategy that matches the first claim verification identification module score and the second claim verification identification module score.

[0035] Optionally, the device further comprises: a training module configured to:

[0036] Obtain sample policy information and build an initialized Light GBM model;

[0037] Get the first and second derivatives of the binary cross entropy function in the Light GBM model;

[0038] The gradient modulus of the sample policy information is calculated based on the first-order derivative and the second-order derivative;

[0039] Transforming the gradient modulus according to a preset density parameter to obtain a target gradient density;

[0040] Constructing a target loss function based on an imbalanced loss function according to the target gradient density;

[0041] The sample policy information is used to train the Light GBM model deployed with the target loss function to obtain a claims settlement strategy perceptron.

[0042] Optionally, the training module is further configured to:

[0043] The target loss function is adjusted by the accumulated value of the sample quantity of the sample policy information.

[0044] Optionally, the training module is further configured to:

[0045] ;

[0046] ;

[0047] Wherein, the GD(g) is the target gradient density, the ( ) is the number of samples in a unit area, is the unit area length, which can be as small as close to 0, g is the gradient modulus of the sample, and x is the number of samples in the unit area.

[0048] Optionally, the training module is further configured to:

[0049] ;

[0050] ;

[0051] Among them, the is the inverse of the gradient density of the i-th sample, N is the total number of samples, is the gradient modulus of the i-th sample, is the unit area length, which can be as small as close to 0, GD(g) is the target gradient density, ( ) is the number of samples in a unit area, is the adjusted objective loss function, the CE ( , ) is the binary cross entropy loss function of the i-th sample, is the sample category probability output by the activation function of the i-th sample, is the category value of the i-th sample.

[0052] Optionally, the processing module is further configured to:

[0053] Combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score;

[0054] When the comprehensive claim identification score is less than the claim identification score threshold, the policy information is submitted to manual claim verification;

[0055] When the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the policy information is submitted to the intelligent claim model for claim settlement.

[0056] Optionally, the training module is further configured to:

[0057] Obtain initial policy information;

[0058] The initial insurance policy is processed according to a target preprocessing method to obtain sample insurance policy information, and the target preprocessing method includes: at least one of: outlier processing, data visualization analysis, and data standardization.

[0059] An embodiment of the present application provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned method for processing insurance policy claims when executed by the processor.

[0060] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for processing insurance policy claims as described above are implemented.

[0061] The embodiments of the present application provide a method, device, electronic device and computer-readable storage medium for processing insurance policy claims. The claims are reviewed by using a claims strategy perception machine improved based on an unbalanced loss function of a claims scenario. The review standards are unified, which reduces the huge demand for human resources in the claims process. While ensuring the accuracy of claims identification, the efficiency of claims is improved. A claims weight coefficient based on a multi-layer perceptron is proposed, which reduces the risk of claims by identifying rejected cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flowchart schematically illustrates a method for processing insurance policy claims in an embodiment of the present application;

[0063] Figure 2 A schematic diagram of a multi-layer perceptron structure in an embodiment of the present application is shown;

[0064] Figure 3A schematic diagram of the claim rejection prediction result of the original model in the embodiment of the present application is schematically shown;

[0065] Figure 4 A schematic diagram of the influence ranking of customer information variables of the original model in an embodiment of the present application is schematically shown;

[0066] Figure 5 A schematic diagram of the prediction result of claim rejection of the optimization model in the embodiment of the present application is schematically shown;

[0067] Figure 6 A schematic diagram schematically shows the influence ranking of customer information variables in the optimization model in the embodiment of the present application;

[0068] Figure 7 The following schematically shows a claim cost analysis diagram provided by an embodiment of the present application;

[0069] Figure 8 A schematic diagram schematically shows a cost comparison diagram of the original model and the optimized model in an embodiment of the present application;

[0070] Fig. 9 The comprehensive claims scoring and processing decision diagram provided in the embodiment of the present application is schematically shown;

[0071] Fig.10 The following is a flowchart schematically showing a method for training a claims settlement strategy perception machine according to an embodiment of the present application;

[0072] Fig.11 The sample gradient modulus distribution statistics diagram provided by the embodiment of the present application is schematically shown;

[0073] Fig.12 The schematic diagram of the distribution of customer groups and claim rejection rates of different products provided in the embodiments of the present application is schematically shown;

[0074] Fig.13 A schematic diagram of the structure of a device for processing insurance policy claims in an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0075] 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.

[0076] Figure 1 A flowchart schematically shows a method for processing insurance policy claims provided in an embodiment of the present application, the method comprising:

[0077] Step 101, obtaining the insurance policy information to be processed.

[0078] In the embodiment of the present application, the policy information refers to the insurance contract that needs to be claimed and the claim proof information provided by the insured. After purchasing the insurance, the insured can submit the claim proof information to the insurance provider for claim settlement after the scenario events specified in the insurance contract occur.

[0079] Step 102, input the policy information into a claims strategy perceptron for processing to obtain a first claims identification module score, and input the policy information into a multi-layer perceptron for processing to obtain a second claims identification module score. The claims strategy perceptron is trained based on a loss function of an imbalanced loss function.

[0080] In the embodiment of the present application, the claim strategy sensor is a machine learning model used to provide a scoring basis for selecting a claim strategy. It should be noted that the claim strategy sensor is not used to verify the insurance policy information, but to provide a score for selecting a claim strategy. The claim strategy sensor is a Light GBM model modified from the loss function of the imbalanced loss function.

[0081] Furthermore, the Light GBM algorithm uses GBDT (Gradient Boosting Decision Tree) as a framework. This framework uses weak classifiers for iterative training to obtain the optimal model, which has the advantages of good training effect and low overfitting. At the same time, Light GBM optimizes the following points based on GBDT and XGBoost: Histogram-based decision tree algorithm; Gradient-based One-Side Sampling (GOSS): GOSS can reduce a large number of data instances with only small gradients, so that when calculating information gain, only the remaining data with high gradients need to be used, which saves a lot of time and space overhead compared to XGBoost traversing all feature values; Exclusive Feature Bundling (EFB): EFB can be used to bind many mutually exclusive features into one feature, thus achieving the purpose of dimensionality reduction; Leaf-wise leaf growth strategy with depth restriction: Most GBDT tools use an inefficient level-wise decision tree growth strategy, because it treats leaves of the same layer indiscriminately, which brings a lot of unnecessary overhead. In fact, the split gain of many leaves is low, so there is no need to search and split. LightGBM uses a leaf-wise algorithm with depth limit; directly supports categorical features; and supports efficient parallel cache hit rate optimization.

[0082] Further, see Figure 2 , the Light GBM after loss function transformation can better train the model, but the efficiency of the claims process needs to be comprehensively evaluated in combination with the labor cost of manual claims. The compensation cost and labor review cost of applying for claims need to be reasonably configured to maximize the benefits. For example, when the claim amount is equivalent to the labor review cost, you can choose to make a direct claim. When the comprehensive index reaches a certain threshold, you can choose to use the intelligent claims model to make a claim rejection judgment. When the comprehensive index reaches a certain threshold, you need to choose manual claims. Therefore, the strategy configuration of claims review is also a starting point for improvement and innovation. Here, it is necessary to build a multi-layer perceptron and introduce a hidden layer. By building a multi-layer perceptron, we can learn whether manual review is required. The input is the variable related to the insured user and the claim amount. The hidden layer and the output layer both use nonlinear activation functions. The output layer is the probability that manual review is required, which is the weight of the claims strategy. Therefore, based on the imbalanced loss function, the loss function is transformed to solve the imbalance problem of positive and negative samples in insurance claims data, the model accuracy is optimized, and a multi-layer perceptron is constructed according to the variables related to the insured user to learn the optimal weight coefficient of the claims processing process configuration, which helps to improve the efficiency of claims.

[0083] Step 103: Perform a claim verification operation on the insurance policy information based on a claim verification review strategy that matches the first claim verification identification module score and the second claim verification identification module score.

[0084] In the embodiment of the present application, the claim review strategy matched by combining the first claim identification module score with the second claim identification module score is constructed according to the variables related to the insured user, and the optimal weight coefficient of the claim processing flow configuration is learned, thereby overcoming the inaccurate review results caused by uneven sample distribution. The claim review strategy can be manually verified, automatically verified by the system, or directly rejected. It can be set according to actual needs and is not limited here.

[0085] For example, the variables of the standardized data set obtained from point 1 to point 3 are entered into the model, and the parameters are set according to the above improved algorithm. The results predicted by the original Light GBM algorithm can be referred to Figure 3 .

[0086] Figure 3 The horizontal axis in the figure, from 1 to 10, is the customer group ranking from low to high probability of claim rejection. The KS curve refers to the Lorenz curve. The larger the KS (Kolmogorov-Smirnov) value, the greater the degree to which the model can distinguish between positive and negative customers (in this experiment, it refers to customers who paid compensation and customers who did not pay compensation). The value range of KS is [0, 1]. Generally speaking, KS>0.2 means that the model has good prediction accuracy. The KS value is as high as 55% in the training set and 49% in the test set. The prediction performance of the original model is already superior enough.

[0087] In addition, the model also outputs the customer variable influence ranking, please refer to Figure 4 :

[0088] Can be obtained from Figure 4 It is found that among all customer information variables, the one that has the greatest impact on user compensation is the difference in days between the reporting date and the policy effectiveness date. In other words, the longer the time interval between the reporting date and the policy effectiveness date, the greater the possibility that the company will pay compensation.

[0089] After inputting the above data set into the improved Light GBM model for claim rejection prediction, the results obtained are as follows: Figure 5 .

[0090] Figure 5 The numbers 1 to 10 are the customer group rankings from low to high probability of claim rejection. Compared with the original model without optimization, the KS value has been significantly improved. The specific comparison is shown in Table 1:

[0091]

[0092] Table 1

[0093] It can be seen that the KS value of the optimized model has been significantly improved, indicating that the efficiency of the optimized model is improving, saving time and resources. In addition, the claims optimization model outputs the customer variable influence ranking for reference Figure 6 .

[0094] Reference Figure 6 The customer information variable with the highest influence ranking obtained by the compensation optimization model is the difference in days between the reporting date and the effective date of the policy. However, the influence of the insured's age has increased, and variables such as the insured's age and premium have appeared that are highly correlated with the insured's personal information, indicating that the influence ranking obtained by the optimized model is more practical.

[0095] Since we use an intelligent claims settlement model to replace part of the manual claims settlement work, we can compare the claims settlement cost of adding the model with the cost of using manual claims settlement alone, and analyze the benefits of using the intelligent claims settlement model.

[0096] This simulation assumes that the original survey method does not investigate policies below 30,000 yuan, so the loss amount of the original method is the amount of compensation for rejected policies in the data. The loss amount of the rejection model is the sum of the correctly predicted policies and the labor investigation cost. In the output results of the original model, the upper limit of the average investigation and labor cost per policy can be referred to as the change of the compensation amount investigation standard Figure 8 .

[0097] from Figure 8 It can be seen from the simulation results that in each lower limit value group of the claim amount investigation, the average investigation and labor costs per order of the optimized model are lower than those of the original model. Therefore, the optimized model and intelligent strategy selection can effectively improve the accuracy and efficiency of claims settlement.

[0098] The embodiment of the present application uses a claims strategy perception machine improved based on the unbalanced loss function of the claims scenario to conduct claims review, unify the review standards, reduce the huge demand for human resources in the claims process, and improve the claims efficiency while ensuring the accuracy of claims identification. A claims weight coefficient based on a multi-layer perceptron is proposed, which reduces the risk of claims by identifying rejected cases.

[0099] Optionally, refer to Fig. 9 , the step 103 may include:

[0100] Step 1031, combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score.

[0101] Step 1032, when the comprehensive claim identification score is less than the claim identification score threshold, the insurance policy information is submitted to manual claim verification.

[0102] Step 1033, when the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the insurance policy information is submitted to the intelligent claim model for claim settlement.

[0103] In the embodiment of the present application, the improved Light GBM model and the output value of the multi-layer perceptron are multiplied to obtain a confidence value, which is defined as formula (1):

[0104] (1)

[0105] in, is the confidence level, is the output value of the Light GBM model, is the output value of the multilayer perceptron.

[0106] The claim identification score threshold is the product of the probabilities corresponding to the turning points of the ROC curves of the two models, which is defined as formula (2):

[0107] (2)

[0108] in, The score threshold for claim identification, is the probability corresponding to the turning point of the ROC curve of the Light GBM model, is the probability corresponding to the turning point of the ROC curve of the multilayer perceptron.

[0109] Therefore, the final claims settlement strategy configuration can be simplified to formula (3):

[0110] (3)

[0111] By improving the Light GBM model and the multi-layer perceptron, we can not only predict customer claim rejections, but also intelligently select claims settlement strategies, which can greatly improve review efficiency.

[0112] Fig.10 A flowchart schematically shows a method for training a claims settlement strategy perception machine according to an embodiment of the present application, the method comprising:

[0113] Step 201, obtain sample policy information and build an initialized Light GBM model.

[0114] Step 202: Obtain the first-order derivative and the second-order derivative of the binary cross entropy function in the Light GBM model.

[0115] Step 203, calculating the gradient modulus of the sample policy information based on the first-order derivative and the second-order derivative.

[0116] Step 204: transform the gradient modulus according to a preset density parameter to obtain a target gradient density.

[0117] Step 205, adjusting the target loss function by the accumulated value of the sample quantity of the sample policy information.

[0118] Step 206: construct a target loss function based on the imbalanced loss function according to the target gradient density.

[0119] Step 207, using the sample policy information to train the Light GBM model deployed with the target loss function to obtain a claims settlement strategy perceptron.

[0120] In the embodiment of step 201 to step 207 of the present application:

[0121] The data sample category is unbalanced, that is, the difference in the amount of positive and negative sample data is too large. The problem of difficulty in distinguishing positive and negative samples can be solved by the cross entropy loss function. The binary cross entropy loss function is as follows:

[0122] (4)

[0123] Among them, CE is the binary cross entropy loss function, It is a parameter that controls the ratio of the positive and negative sample loss functions, p is the sample category probability output by the sigmoid function, y is the category value of the sample, and the sigmoid function is an activation function.

[0124] However, according to actual insurance claims data, there is not only a problem of too large a difference in the number of positive and negative samples. In the process of insurance signing and claims, information asymmetry such as moral hazard frequently occurs, which is reflected in the claims data as difficult samples, especially difficult negative samples. In order to solve the problem of identifying difficult and easy samples, we can start from the confidence level to reduce the loss and obtain FocalLoss, as shown in the following formula (5):

[0125] (5)

[0126] Among them, FL is the loss function Focal Loss, p is the sample category probability output by the sigmoid function, y is the category value of the sample, is a parameter that controls the class probability ratio.

[0127] However, Focal Loss cannot solve the problem of over-focusing on difficult-to-classify samples. Therefore, this technology uses an improved imbalanced loss function to solve the problem of imbalanced binary classification samples.

[0128] In order to control the loss size of each sample, it is necessary to find the gradient modulus of each sample, that is, to find the partial derivative of the original binary cross entropy loss function with respect to x to obtain the following formula (6):

[0129] (6)

[0130] Therefore, after taking the partial derivative and combining the y values ​​in the two cases, we can get the gradient modulus of each sample, which is the following formula (7):

[0131] (7)

[0132] in, is the gradient modulus of the i-th sample, p is the sample category probability output by the sigmoid function, and y is the category value of the sample.

[0133] Reference Fig.11 , g has a larger sample distribution ratio at both boundaries (0, 1), with samples on the left being easy to separate and samples on the right being difficult to separate, i.e., some outliers. It is necessary to define a parameter so that after transformation, the samples at both boundaries are less sensitive to the loss function. Therefore, the gradient density can be defined, i.e., the ratio of the number of samples in a unit area to the length of the unit area, as shown in the following formula (8):

[0134] (8)

[0135] Where CD(g) is the target gradient density, ( ) is the number of samples in the unit area, is the unit area length and can be as small as close to 0.

[0136] The gradient modulus of the i-th sample is about The number of samples contained in the unit area length of the interval, that is, the number of samples in the unit area, can be expressed as the following formula (9):

[0137] (9)

[0138] You can define a parameter It represents the inverse of the sample gradient density, which is used to suppress the contribution of outliers with higher sample gradient modulus, as shown in the following formula (10):

[0139] (10)

[0140] In addition to using gradient density to control sample weights, the numerator also refines the degree of suppression of easily separable samples and outlier samples by limiting the number of samples accumulated.

[0141] The improved loss function is It can be expressed as the following formula (11)

[0142] (11)

[0143] Among them, CE ( , ) represents the binary cross entropy loss function of the i-th sample. For details, please refer to the above formula (4). GD ( ) represents the ratio of the number of samples in the unit area of ​​the i-th sample to the length of the unit area. For details, please refer to the above formula (8), and N is the total number of samples.

[0144] Based on the derivation process of some of the above loss functions, implementing a custom loss function in Light GBM requires the following steps:

[0145] A1. Write out the constructed loss function;

[0146] A2. Write the first and second derivatives of the loss function;

[0147] A3. Rewrite the training loss function and evaluation function respectively, which are the parameters fobj and feval in the Light GBM interface;

[0148] A4. After customizing the loss function, the original initial value parameters become invalid and you need to manually assign the sample initial value parameters.

[0149] A5. The predictions obtained by the custom loss function need to be manually converted to the sigmoid function.

[0150] Optionally, the step 201 may include:

[0151] B1, obtain initial policy information.

[0152] In the embodiment of the present application, since the dimensions of insurance customer data are high and there is duplicate information between different tables, it is necessary to screen and merge the data, select reasonable variables for model entry, delete redundant information, organize it using database technology and save it into a table, annotate it with claims labels, and enter the supervised learning model to generate training sets, validation sets and test sets for claim rejection analysis to obtain initial policy information.

[0153] B2. Processing the initial insurance policy according to a target preprocessing method to obtain sample insurance policy information. The target preprocessing method includes: at least one of outlier processing, data visualization analysis, and data standardization.

[0154] In the embodiment of the present application, the missing value processing is because different salesmen have different ways of entering customer information, so there may be duplicate labels and missing customer information. If the missing values ​​are not processed, the data will become sparse, affecting the effect of the final data modeling. During the data cleaning process, statistics are required for the data under the same label. If the missing value exceeds 80%, then this label is regarded as an invalid label and deleted.

[0155] Outliers are data that deviate from the majority of data in a data set. In terms of data values, data that deviate from the mean by more than two standard deviations, and data that deviate from the mean by more than three standard deviations, are called highly abnormal outliers. When the amount of sample data is scarce, outliers can be replaced with the mean to minimize the impact of outliers while keeping the amount of data unchanged.

[0156] Data visualization analysis is to check the distribution characteristics of the data set before the insurance claim data enters the model, so as to train the model more targeted. This technical solution decided to select the claim data of customers who purchased the company's new products with higher value, and regarded the data with non-normal claims (including rejection, accommodative claims, and negotiated claims) as samples of policies that should be rejected. Fig.12 shown.

[0157] from Fig.12 It can be seen that different insurance products differ in terms of quantity share and claim rejection rate. At the same time, due to the characteristics of insurance products, the proportion of people who have suffered losses to the total number of insured people is relatively low. From a data perspective, there is an imbalance. Therefore, how to perform small sample identification and prediction under the condition of sample imbalance is a daunting challenge.

[0158] Data standardization is due to the differences in the dimensions of each feature in the data, so the original data needs to be standardized. For continuous data, Min-Max, Z-score and other methods can be used for standardization. For categorical variables, dummy variable processing or One-Hot encoding can be performed, thereby summarizing the statistical distribution of a unified sample.

[0159] Fig.13 The structure diagram of a device 40 for processing insurance policy claims provided in an embodiment of the present application is schematically shown, and the device comprises:

[0160] The receiving module 401 is configured to obtain the insurance policy information to be processed;

[0161] The preheating module 402 is configured to input the policy information into a claims settlement strategy perceptron for processing to obtain a first claims settlement identification module score, and input the policy information into a multilayer perceptron for processing to obtain a second claims settlement identification module score, wherein the claims settlement strategy perceptron is trained based on a loss function of an imbalanced loss function;

[0162] The processing module 403 is configured to perform a claim verification operation on the insurance policy information based on a claim verification review strategy that matches the first claim verification identification module score and the second claim verification identification module score.

[0163] Optionally, the device further comprises: a training module configured to:

[0164] Obtain sample policy information and build an initialized Light GBM model;

[0165] Get the first and second derivatives of the binary cross entropy function in the Light GBM model;

[0166] The gradient modulus of the sample policy information is calculated based on the first-order derivative and the second-order derivative;

[0167] Transforming the gradient modulus according to a preset density parameter to obtain a target gradient density;

[0168] Constructing a target loss function based on an imbalanced loss function according to the target gradient density;

[0169] The sample policy information is used to train the Light GBM model deployed with the target loss function to obtain a claims settlement strategy perceptron.

[0170] Optionally, the training module is further configured to:

[0171] The target loss function is adjusted by the accumulated value of the sample quantity of the sample policy information.

[0172] Optionally, the training module is further configured to:

[0173] ;

[0174] ;

[0175] Wherein, the GD(g) is the target gradient density, the ( ) is the number of samples in a unit area, is the unit area length, which can be as small as close to 0, g is the gradient modulus of the sample, and x is the number of samples in the unit area.

[0176] Optionally, the training module is further configured to:

[0177] ;

[0178] ;

[0179] Among them, the is the inverse of the gradient density of the i-th sample, N is the total number of samples, is the gradient modulus of the i-th sample, is the unit area length, which can be as small as close to 0, GD(g) is the target gradient density, ( ) is the number of samples in a unit area, is the adjusted objective loss function, the CE ( , ) is the binary cross entropy loss function of the i-th sample, is the sample category probability output by the activation function of the i-th sample, is the category value of the i-th sample.

[0180] Optionally, the processing module 403 is further configured to:

[0181] Combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score;

[0182] When the comprehensive claim identification score is less than the claim identification score threshold, the policy information is submitted to manual claim verification;

[0183] When the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the policy information is submitted to the intelligent claim model for claim settlement.

[0184] Optionally, the training module is further configured to:

[0185] Obtain initial policy information;

[0186] The initial insurance policy is processed according to a target preprocessing method to obtain sample insurance policy information, and the target preprocessing method includes: at least one of: outlier processing, data visualization analysis, and data standardization.

[0187] The embodiment of the present application uses a claims strategy perception machine improved based on the unbalanced loss function of the claims scenario to conduct claims review, unify the review standards, reduce the huge demand for human resources in the claims process, and improve the claims efficiency while ensuring the accuracy of claims identification. A claims weight coefficient based on a multi-layer perceptron is proposed, which reduces the risk of claims by identifying rejected cases.

[0188] An embodiment of the present application provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned method for processing insurance policy claims when executed by the processor.

[0189] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for processing insurance policy claims as described above are implemented.

[0190] Optionally, an embodiment of the present application also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the various processes of the above-mentioned information processing method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0191] The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, each process of the above-mentioned information processing method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0192] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device 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 device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0193] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0194] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

[0195] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0196] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0197] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0198] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0199] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0200] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.

[0201] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for processing insurance claims, characterized in that: The method comprises: Get pending policy information; Input the policy information into the claims strategy perceptron for processing to obtain a first claims identification module score, and input the policy information into the multilayer perceptron for processing to obtain a second claims identification module score, wherein the claims strategy perceptron is obtained by loss function training based on an imbalanced loss function, and the claims strategy perceptron is used to provide a score for selecting a claims strategy, and the output layer of the multilayer perceptron outputs the probability of requiring manual claims; Performing a claim verification operation on the policy information based on a claim verification review strategy in which the first claim verification identification module score and the second claim verification identification module score match; The claim verification strategy based on the matching of the first claim verification identification module score and the second claim verification identification module score performs a claim verification operation on the policy information, including: Combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score; When the comprehensive claim identification score is less than the claim identification score threshold, the policy information is submitted to manual claim verification; When the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the policy information is submitted to the intelligent claim model for claim settlement.

2. The method according to claim 1, characterized in that The claim settlement strategy perceptron is trained through the following steps: Obtain sample policy information and build an initialized Light GBM model; Get the first and second derivatives of the binary cross entropy function in the Light GBM model; The gradient modulus of the sample policy information is calculated based on the first-order derivative and the second-order derivative; Transforming the gradient modulus according to a preset density parameter to obtain a target gradient density; Constructing a target loss function based on an imbalanced loss function according to the target gradient density; The sample policy information is used to train the Light GBM model deployed with the target loss function to obtain a claims settlement strategy perceptron.

3. The method according to claim 2, characterized in that After constructing a target loss function based on an imbalanced loss function according to the target gradient density, the method includes: The target loss function is adjusted by the accumulated value of the sample quantity of the sample policy information.

4. The method according to claim 2, characterized in that: The step of transforming the gradient modulus according to a preset density parameter to obtain a target gradient density includes: ； ； Wherein, the GD(g) is the target gradient density, the ( ) is the number of samples in a unit area, is the unit area length, which can be as small as close to 0, g is the gradient modulus of the sample, and x is the number of samples in the unit area.

5. The method according to claim 3, characterized in that: The adjusting the target loss function by the accumulated value of the sample quantity of the sample policy information includes: ； ； Among them, the is the inverse of the gradient density of the i-th sample, N is the total number of samples, is the gradient modulus of the i-th sample, is the unit area length, which can be as small as close to 0, GD(g) is the target gradient density, ( ) is the number of samples in a unit area, is the adjusted objective loss function, the CE ( , ) is the binary cross entropy loss function of the i-th sample, is the sample category probability output by the activation function of the i-th sample, is the category value of the i-th sample.

6. The method according to claim 2, characterized in that The obtaining of sample policy information includes: Obtain initial policy information; The initial insurance policy is processed according to a target preprocessing method to obtain sample insurance policy information, and the target preprocessing method includes: at least one of: outlier processing, data visualization analysis, and data standardization.

7. A claim verification device for insurance policy settlement, characterized in that: The device comprises: A receiving module, configured to obtain insurance policy information to be processed; The prediction module is configured to input the policy information into a claims settlement strategy perceptron for processing to obtain a score for the first claims settlement identification module, and input the policy information into a multilayer perceptron for processing to obtain a score for the second claims settlement identification module, wherein the claims settlement strategy perceptron is obtained by training a loss function based on an imbalanced loss function, the claims settlement strategy perceptron is used to provide a score for selecting a claims settlement strategy, and the output layer of the multilayer perceptron outputs a probability of requiring manual claims settlement; A processing module configured to perform a claim verification operation on the policy information based on a claim verification review strategy in which the first claim verification identification module score and the second claim verification identification module score match; The processing module is further configured to: Combining the first claim identification module score and the second claim identification module score to obtain a comprehensive claim identification score; When the comprehensive claim identification score is less than the claim identification score threshold, the policy information is submitted to manual claim verification; When the comprehensive claim identification score is greater than or equal to the claim identification score threshold, the policy information is submitted to the intelligent claim model for claim settlement.

8. An electronic device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of a method for processing insurance policy claims as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for processing insurance policy claims as described in any one of claims 1 to 6.

Images