Group accidental insurance fraud detection method and system based on abnormal insurance

Abstract

The invention provides an abnormal insurance behavior identification method and system based on a legal person business knowledge graph, and the method comprises the steps: screening associated risk feature dimensions according to the correlation with the loss ratio, and determining a risk threshold value corresponding to each risk feature dimension according to the screened risk feature dimensions; calculating a risk feature value corresponding to each risk feature dimension of the to-be-identified legal person service insurance policy, and screening the legal person service insurance policy of which the risk feature value is higher than a risk threshold value according to the risk feature value corresponding to each risk feature dimension of the to-be-identified legal person service insurance policy and the risk threshold value corresponding to each risk feature dimension; determining an expected loss ratio corresponding to the legal person business insurance policy through a pre-established loss ratio model and the risk feature value corresponding to each risk feature dimension; and outputting the legal person service insurance policy of which the expected compensation ratio is higher than the preset compensation ratio threshold. The timeliness and the reliability of group accidental insurance fraud detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of insurance fraud detection and prevention, and in particular to a group accident insurance fraud detection method and system based on abnormal insurance. Background Art

[0002] Group accident insurance fraud is generally a rare event, but it causes great losses to insurance companies.

[0003] At present, group accident insurance anti-fraud is mostly carried out in the claims stage, using internal and external big data combined with artificial intelligence algorithms to conduct quantitative or qualitative risk assessments on the insured who report the claim, and then determine the possibility of insurance fraud. However, given the difficulty in investigating and collecting evidence in some claims cases, such as accidents in other places, fraud monitoring in the claims stage often does not play its due role. Therefore, fraud monitoring is pre-positioned in the insurance stage, and the probability of insurance fraud is calculated using algorithms. Accident insurance group insurance cases with a higher probability can be handed over to the underwriter or company investigation post for risk investigation, and then measures such as increasing premiums or conditional underwriting can be taken, so that insurance fraud detection and prevention can be achieved at the entry point.

[0004] However, group accident insurance fraud is generally a rare event, lacking a large number of effective training samples, and there is no fraud detection for group accident insurance at the time of insurance, which makes the timeliness and reliability of group accident insurance fraud detection low.

[0005] In summary, the timeliness and reliability of fraud detection in group accident insurance in the prior art are not high. In order to solve the problem of low timeliness and reliability of fraud detection in group accident insurance in the prior art, this solution is proposed. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the present invention innovatively proposes a group accident insurance fraud detection method and system based on abnormal insurance, which effectively solves the problems of low timeliness and reliability of fraud detection in group accident insurance caused by the prior art, and effectively improves the timeliness and reliability of fraud detection in group accident insurance.

[0007] The first aspect of the present invention provides a group accident insurance fraud detection method based on abnormal insurance, comprising:

[0008] According to the correlation with the loss ratio, the associated risk characteristic dimensions are screened, and the risk threshold corresponding to each risk characteristic dimension is determined based on the screened risk characteristic dimensions; among which, the screened associated risk characteristic dimensions include the length of legal person registration, the frequency of adding people, and the proportion of non-local insured persons;

[0009] Obtain the corporate business insurance policies to be identified, calculate the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified, and screen corporate business insurance policies whose risk characteristic values ​​are higher than the risk threshold according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified and the risk threshold corresponding to each risk characteristic dimension;

[0010] For corporate business policies with risk characteristic values ​​higher than the risk threshold, the expected loss ratio of the corporate business policy is determined through the pre-established loss ratio model and the risk characteristic values ​​corresponding to each risk characteristic dimension;

[0011] Output corporate business policies whose expected loss ratio is higher than the preset loss ratio threshold.

[0012] A second aspect of the present invention provides a group accident insurance fraud detection system based on abnormal insurance, comprising:

[0013] The determination module selects the associated risk characteristic dimensions according to the correlation with the loss ratio, and determines the risk threshold corresponding to each risk characteristic dimension according to the selected risk characteristic dimensions; wherein the associated risk characteristic dimensions after screening include the registration time of the legal person, the frequency of adding people, and the proportion of non-local insured persons;

[0014] The screening module obtains the corporate business insurance policies to be identified, calculates the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified, and screens the corporate business insurance policies whose risk characteristic values ​​are higher than the risk threshold according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified and the risk threshold corresponding to each risk characteristic dimension;

[0015] The calculation module determines the expected loss ratio of the corporate business insurance policy for which the risk characteristic value is higher than the risk threshold through the pre-established loss ratio model and the risk characteristic value corresponding to each risk characteristic dimension;

[0016] The output module outputs corporate business policies whose expected claims ratio is higher than the preset claims ratio threshold.

[0017] The technical solution adopted by the present invention includes the following technical effects:

[0018] 1. In the technical solution of the present invention, the associated risk feature dimensions are screened according to the correlation with the claims ratio, and the risk threshold corresponding to each risk feature dimension is determined according to the screened risk feature dimensions; the corporate business policy to be identified is obtained, and the risk feature value corresponding to each risk feature dimension of the corporate business policy to be identified is calculated; according to the risk feature value corresponding to each risk feature dimension of the corporate business policy to be identified and the risk threshold corresponding to each risk feature dimension, the corporate business policies with risk feature values ​​higher than the risk threshold are screened; for the corporate business policies with risk feature values ​​higher than the risk threshold, the expected claims ratio corresponding to the corporate business policy is determined by a pre-established claims ratio model and the risk feature value corresponding to each risk feature dimension; the corporate business policies with expected claims ratio higher than the preset claims ratio threshold are output, so as to effectively solve the problem of low timeliness and reliability of fraud detection of group accident insurance caused by the prior art, and effectively improve the timeliness and reliability of fraud detection of group accident insurance.

[0019] 2. The technical solution of the present invention is based on group accident insurance policies in the past three years to establish a data set consisting of feature dimensions and claims ratios; calculate the covariance between each feature dimension and the claims ratio, and retain all feature dimensions corresponding to non-zero covariances as risk feature dimensions; calculate the correlation coefficient between each risk feature dimension and the claims ratio; sort by the absolute value of the correlation coefficient, and screen out several risk feature dimensions with the highest correlation as associated risk feature dimensions, thereby ensuring the reliability of fraud detection in group accident insurance.

[0020] 3. The technical solution of the present invention screens out all the insured persons whose legal person registration time of the insured enterprise is within two years at the time of insurance, and calculates the simple claims ratio of each screened insured person in the past three years; determines the monthly registration value of the insured whose simple claims ratio exceeds 100%, and establishes a legal person registration time data set, an added person frequency data set, and a non-local insured person proportion data set; based on the quantile method or standard deviation method of the data in the legal person registration time data set, the added person frequency data set, and the non-local insured person proportion data set, the risk threshold corresponding to each feature dimension is determined, and traditional statistical methods can be used to perform statistical analysis on limited sample statistical data to determine the risk threshold of the associated risk feature dimension, thereby ensuring the reliability of fraud detection in group accident insurance.

[0021] 4. In the technical solution of the present invention, based on the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified and the risk threshold value corresponding to each risk characteristic dimension, screening corporate business policies with risk characteristic values ​​higher than the risk threshold value specifically includes: screening out corporate business policies whose risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified are all higher than the risk threshold value corresponding to each risk characteristic dimension; or, screening out corporate business policies whose any one of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified is higher than the risk threshold value corresponding to the risk characteristic dimension; or, screening out corporate business policies whose weighted risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified is higher than the risk weighted threshold value corresponding to the weighted risk characteristic dimension. According to risk requirements, corporate business policies with risk characteristic values ​​higher than the risk threshold value can be flexibly selected, thereby improving the scope of application of fraud detection in group accident insurance.

[0022] 5. In the technical solution of the present invention, the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is obtained, and the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is input into a pre-established loss ratio decision tree model, or a loss ratio logistic regression model, or a loss ratio regression model; the total loss ratio output by the pre-established loss ratio decision tree model, or the abnormal insurance probability output by the loss ratio logistic regression model, or the total loss ratio output by the loss ratio regression model is used as the expected loss ratio of the corporate business policy. The loss ratio model can be flexibly selected according to actual requirements, thereby improving the scope of application and flexibility of fraud detection in group accident insurance.

[0023] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is a schematic diagram of a process of a method in Example 1 of the present invention;

[0026] Figure 2 This is a partial flow diagram of step S1 of the method in Example 1 of the present invention;

[0027] Figure 3 This is another partial flow diagram of step S1 of the method in Example 1 of the present invention;

[0028] Figure 4 Schematic diagram of the process of step S2 in the method of embodiment 1 of the present invention;

[0029] Figure 5 It is a schematic diagram of the structure of the system of Example 2 in the solution of the present invention. DETAILED DESCRIPTION

[0030] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

[0031] Embodiment 1

[0032] like Figure 1 As shown, the present invention provides a group accident insurance fraud detection method based on abnormal insurance, comprising:

[0033] S1, based on the correlation with the loss ratio, screen the associated risk characteristic dimensions, and determine the risk threshold corresponding to each risk characteristic dimension based on the screened risk characteristic dimensions; wherein the screened associated risk characteristic dimensions include the length of legal person registration, the frequency of adding people, and the proportion of non-local insured persons;

[0034] S2, obtaining the corporate business insurance policies to be identified, calculating the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified, and screening corporate business insurance policies whose risk characteristic values ​​are higher than the risk threshold according to the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified and the risk threshold corresponding to each risk characteristic dimension;

[0035] S3, for corporate business insurance policies with risk characteristic values ​​higher than the risk threshold, the expected loss ratio corresponding to the corporate business insurance policy is determined through the pre-established loss ratio model and the risk characteristic values ​​corresponding to each risk characteristic dimension;

[0036] S4, output corporate business policies whose expected loss ratio is higher than the preset loss ratio threshold.

[0037] In this embodiment, group accident insurance fraud is generally a rare event. Therefore, if artificial intelligence algorithms such as machine learning are used, it is almost impossible to construct a large number of effective training samples. Therefore, methods based on traditional statistics are more suitable.

[0038] Among them, Figure 2 As shown, in step S1, according to the correlation with the loss ratio, the associated risk feature dimensions are screened, specifically including:

[0039] S101, based on the group accident insurance policies in the past three years, establish a data set consisting of feature dimensions and claim ratios;

[0040] First, dozens of feature dimensions were formed based on the policyholder, the insured, the static attributes of the policy, and the dynamic attributes of the policy. Then, based on the group accident insurance policies in the past three years, these risk dimensions and claims ratios were combined into a data set.

[0041] S102, calculating the covariance between each feature dimension and the loss ratio, and retaining all feature dimensions corresponding to non-zero covariances as risk feature dimensions;

[0042] The feature dimension data and loss ratio data in the data set are constructed based on the feature dimensions and loss ratio. The covariance between each feature dimension and the loss ratio is calculated according to the following formula. The features with a covariance of 0 are eliminated, and all non-0 feature dimensions are retained. The feature dimensions related to the loss ratio are screened out as risk feature dimensions.

[0043] Among them, X is the feature dimension data (sample data) in the data set, Y is the loss ratio data (sample data) in the data set, X i is the i-th feature dimension data in the data set, Y i is the loss ratio data corresponding to the i-th feature dimension data in the data set, n is the number of feature dimension data in the data set, is the mean of the i-th feature dimension data in the data set, It is the mean of the loss ratio data corresponding to the i-th feature dimension data in the data set.

[0044] S103, calculating the correlation coefficient between each risk characteristic dimension and the loss ratio; wherein the correlation coefficient is the ratio of a first calculation factor to a second calculation factor, wherein the first calculation factor is the covariance between each risk characteristic dimension and the loss ratio, and the second calculation factor is the standard deviation between each risk characteristic dimension and the loss ratio;

[0045] The correlation coefficient is calculated as:

[0046] Among them, r XY represents the correlation coefficient between the sample data of each risk characteristic dimension and the sample data of the loss ratio, S XY represents the covariance of the risk characteristic dimension sample data and the loss ratio sample data, S Xrepresents the sample standard deviation of each risk feature dimension sample data X (feature dimension sample data), S Y It represents the sample standard deviation of the sample data Y of the loss ratio. The following are S XY Covariance and S X and S Y The formula for calculating standard deviation.

[0047]

[0048] S X Sample standard deviation calculation formula:

[0049]

[0050] S Y Sample standard deviation calculation formula:

[0051]

[0052] The covariance and standard deviation of the sample data of each risk characteristic dimension and the sample data of the loss ratio were calculated respectively, and the correlation coefficient value was obtained.

[0053] S104, sorting by absolute value of the correlation coefficient, and selecting several risk feature dimensions with the highest correlation as associated risk feature dimensions.

[0054] The absolute values ​​of the correlation coefficients are sorted from large to small, and the three risk dimensions with the highest correlation are screened out: the length of legal person registration, the frequency of adding people, and the proportion of insured persons in other places.

[0055] Among them, the number of risk feature dimensions can be 3, and can also be flexibly adjusted according to actual needs, and the present invention does not limit this.

[0056] Among them, Figure 3 As shown, in step S1, determining the risk threshold corresponding to each risk feature dimension according to the screening risk feature dimension includes:

[0057] S111, screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years; among them, the insured person of group accident insurance is the legal person insured;

[0058] Simple claim ratio refers to the percentage of claim expenditure in a period of time to premium income. The specific calculation formula is: Simple claim ratio = (settled claims + pending claims) / premium income × 100%. The insured of group accident insurance is a legal person, that is, the insured of group accident insurance is a corporate legal person.

[0059] S112, determining the monthly registration value of the insured whose simple claim ratio exceeds 100%, and establishing a legal person registration duration data set; determining the risk threshold corresponding to the legal person registration duration based on the quantile method or standard deviation method of the legal person registration duration data of the legal person registration duration data set;

[0060] Legal person registration time: threshold1, defined as the threshold of "legal person registration time". First, screen out the insured whose insured enterprise registration time is less than two years at the time of insurance, then convert the insured enterprise registration into monthly, and then calculate the simple claim ratio of each insured for three years, and finally find the monthly registration value of the insured with a claim ratio exceeding 100% as the data set of "legal person registration time". Secondly, perform descriptive statistical analysis on the data of legal person registration time, including mean, standard deviation, minimum and maximum values.

[0061] The risk threshold corresponding to the legal person registration time is: the 25th percentile of the legal person registration time data in the legal person registration time data set, or the difference between the mean of the legal person registration time data in the legal person registration time data set and the geometric multiple standard deviation.

[0062] That is, (i) quantile method:

[0063] Calculate the quantiles of the legal person registration time (such as the 25th percentile). Set the threshold to the 25th percentile: If the legal person registration time of an insured person is lower than the risk threshold corresponding to the legal person registration time, it is marked as high risk.

[0064] (ii) Standard Deviation Method:

[0065] According to the mean (μ1) and geometric standard deviation (σ1) of the legal person registration time data set, the risk threshold corresponding to the legal person registration time is set as:

[0066] threshold1=μ1-2σ1,

[0067] If the legal entity registration time is lower than this threshold, it is marked as high risk.

[0068] S113, screening out the insurance policies of the insured persons whose simple claim ratio exceeds 100%, calculating the ratio of the number of persons added or reduced in the insurance policies after screening to the total number of persons insured during the effective period of the insurance policies, and establishing a frequency data set of persons added; determining the risk threshold corresponding to the frequency of persons added based on the quantile method or standard deviation method of the frequency data of persons added in the frequency data set of persons added;

[0069] Frequency of increase: threshold2 is defined as the threshold of "frequency of increase". First, calculate the simple claim ratio of each corporate policyholder in the past three years, then filter out the policies of customers with claim ratios exceeding 100%, and then calculate the ratio of the number of increase or decrease in these policies during the validity period relative to the total number of insured persons, as the data set of "frequency of increase". Secondly, calculate the mean, standard deviation, minimum and maximum value of the frequency of increase.

[0070] The risk threshold corresponding to the increase in person frequency is: the 75th percentile of the increase in person frequency data of the increase in person frequency data set, or the sum of the mean and geometric multiple standard deviation of the legal person registration duration data of the legal person registration duration data set.

[0071] That is, (i) quantile method:

[0072] Calculate the 75th percentile of the frequency of added people.

[0073] The risk threshold corresponding to the frequency of increase of persons is set to the 75th percentile of the frequency of increase of persons in the frequency of increase of persons data set: if the frequency of increase of persons of a certain insured person is higher than the risk threshold corresponding to the frequency of increase of persons, the insured person is marked as high risk.

[0074] (ii) Standard Deviation Method:

[0075] According to the mean (μ2) and geometric standard deviation (σ2) of the increased frequency data set, the risk threshold corresponding to the increased frequency is set as:

[0076] threshold2=μ2+2σ2

[0077] If the frequency of adding people is higher than this threshold, it is marked as high risk.

[0078] S114, calculate the ratio of all non-local insured persons to the total number of insured persons during the validity period of the policy after screening, and establish a non-local insured person ratio data set; determine the risk threshold corresponding to the non-local insured person ratio based on the quantile method or standard deviation method of the non-local insured person ratio data of the non-local insured person ratio data set; wherein, the non-local insured person is the insured person whose registered address is not in the same prefecture-level city as the insured enterprise of the group accident insurance.

[0079] Proportion of non-local insured persons: threshold3, defined as the threshold of "proportion of non-local insured persons". First, the simple claims ratio of each corporate policyholder in the past three years is calculated, and then the policies of customers with claims ratios exceeding 100% are screened out. Then, the ratio of all non-local insured persons (i.e. non-local insured persons, insured persons whose registered address is not in the same city as the underwriting company of group accident insurance, determined by ID number) in the validity period of these policies relative to the total number of insured persons is calculated as the data set of "proportion of non-local insured persons". Secondly, a descriptive statistical analysis is performed on the proportion of non-local insured persons, including the mean, standard deviation, minimum and maximum values.

[0080] The risk threshold corresponding to the proportion of non-local insured persons is: the 90th percentile of the proportion of non-local insured persons in the non-local insured persons proportion data set, or the sum of the mean and geometric multiple standard deviation of the proportion of non-local insured persons in the non-local insured persons proportion data set;

[0081] Specifically, (i) Quantile method:

[0082] Calculate the 90th percentile of the proportion of out-of-town insured persons.

[0083] The risk threshold corresponding to the proportion of non-local insured persons is set to the 90th percentile of the non-local insured persons proportion data in the non-local insured persons proportion data set: if the frequency of increase of a certain insured person is higher than the risk threshold corresponding to the non-local insured persons proportion, it is marked as high risk.

[0084] (ii) Standard Deviation Method:

[0085] According to the mean (μ3) and geometric standard deviation (σ3) of the proportion of insured persons in different places in the data set, the risk threshold corresponding to the increase in the frequency of people is set as follows:

[0086] threshold3=μ3+2σ3

[0087] If the proportion of out-of-town insured persons is higher than this threshold, it will be marked as high risk.

[0088] Among them, Figure 4 As shown, in step S2, obtaining the insurance policy of the legal person business to be identified, and calculating the risk characteristic value corresponding to each risk characteristic dimension of the insurance policy of the legal person business to be identified specifically include:

[0089] S201, obtaining a data set of insured information in the insurance policy of the legal person business to be identified, the insured information includes the insurance date, the registration date of the insured legal person, the date of adding or removing the insured, and the insured ID number;

[0090] The policy input parameters for the legal person business to be identified contain a data set of insured information, with the following fields:

[0091] Insurance date: the specific date of insurance purchase;

[0092] Insured legal person registration date: the date of legal person registration, used to calculate registration duration;

[0093] Date of adding or reducing the insured person: record the time when the insured person is added or reduced;

[0094] Insured ID number: The ID number of each insured person, used for unique identification.

[0095] Contract ID: A unique identifier for each insurance contract.

[0096] threshold1, threshold2, threshold3: risk thresholds corresponding to the length of legal person registration, the frequency of adding people, and the proportion of non-local insured persons.

[0097] time_period: The time range for analysis, such as "month" or "quarter", used to determine the period for which the threshold is applied (the risk threshold in the above text does not change accordingly).

[0098] S202, calculating the risk characteristic value corresponding to the legal person registration time according to the insurance date and the insured legal person registration date;

[0099] The risk characteristic value corresponding to the legal person registration duration is the difference between the insurance date and the insured legal person registration date, converted into monthly duration, that is, the legal person registration duration of the insured as of the insurance date.

[0100] S203, calculating the risk characteristic value corresponding to the frequency of adding people after the insurance date but before the insurance takes effect according to the insurance date and the date of adding or reducing the insured person;

[0101] The risk characteristic value corresponding to the frequency of adding people is the frequency of adding, reducing or changing the insured person after the insurance date but before the insurance takes effect, that is, the length of time the insured person has been registered as a legal person as of the insurance date. Specifically, to calculate the number of times the insured person is added or reduced for each insurance contract during the statistical period, you can use the GROUP BY and COUNT functions of SQL, or the groupby and size methods in the pandas library of Python.

[0102] S204, calculating the risk characteristic value corresponding to the proportion of non-local insured persons based on the insured person's ID number.

[0103] The risk characteristic value corresponding to the proportion of out-of-town insured persons is the proportion of out-of-town insured persons to all insured persons before the insurance takes effect after the insurance date, that is, the proportion of out-of-town insured persons before the insurance takes effect after the insurance date.

[0104] Among them, in step S2, according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policy to be identified and the risk threshold corresponding to each risk characteristic dimension, the corporate business insurance policies with risk characteristic values ​​higher than the risk threshold are screened, specifically including the following situations:

[0105] The first type is to screen out the corporate business insurance policies to be identified whose risk characteristic values ​​corresponding to each risk characteristic dimension are higher than the risk threshold corresponding to each risk characteristic dimension;

[0106] Or, the second method is to screen out any of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified that is higher than the risk threshold corresponding to the risk characteristic dimension;

[0107] Or, the third type is to screen out the corporate business policies to be identified whose corresponding risk characteristic weighted values ​​after weighting each risk characteristic dimension are higher than the corresponding risk weighted threshold after weighting the risk characteristic dimension.

[0108] Overall risk assessment: After setting the corresponding risk threshold for each risk characteristic dimension, a comprehensive scoring model can be developed by combining the risk assessment results of these three risk characteristic dimensions:

[0109] For each risk characteristic dimension, a weight is set (such as 0.4, 0.3, 0.3), and a weighted score is performed according to the risk level of each risk characteristic dimension, and finally a comprehensive risk score for each policyholder is obtained. This method can be used as an auxiliary means to help insurance companies effectively identify and screen potential fraud and reduce the risk of loss.

[0110] Preferably, the fourth type may also be included, which is to screen out corporate business policies for which any two of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policies to be identified are higher than the risk threshold corresponding to the risk characteristic dimension.

[0111] In the actual operation process, you can flexibly select four situations according to the actual risk control requirements.

[0112] Among them, in step S3, the pre-established loss ratio model includes a loss ratio regression model, a loss ratio logistic regression model, and a loss ratio decision tree model.

[0113] Specifically, the process of establishing the loss ratio regression model includes:

[0114] Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance application, calculate the simple claim ratio of each screened insured person in the past three years (natural years), the total claim ratio of each screened insured person in the past three years (fiscal years), and the risk characteristic value corresponding to each risk characteristic dimension;

[0115] Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total loss ratio (total loss ratio of each corporate business), legal person registration time (corresponding risk characteristic value), frequency of adding people (corresponding risk characteristic value), and proportion of non-local insured persons (corresponding risk characteristic value);

[0116] The total loss ratio refers to the percentage of loss expenditure to premium income in a certain accounting period. The calculation formula is: loss ratio = (loss expenditure ÷ premium income) × 100%.

[0117] Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of dimensions and ensure the effectiveness of the regression model;

[0118] Taking the total loss ratio as the dependent variable and the risk characteristic value corresponding to each risk characteristic dimension as the independent variable, a multiple linear regression model of the loss ratio (loss ratio regression model) is established;

[0119] The form of the multiple linear regression model of loss ratio is as follows:

[0120] Total Claim Rate=β 10 +β 11 THERESHOLD1+β 12 THERESHOLD2+β 13 ·

[0121] THERESHOLD3+∈1,

[0122] Among them: Total Claim Rate is the total claim rate; β 10 is the intercept term; β 11 , β 12 , β 13 is the regression coefficient of each variable, indicating the degree of influence of each risk characteristic value on the total claims ratio; ∈1 is the error term; THERESHOLD1 is the risk characteristic value corresponding to the length of legal person registration, that is, the variable value of the length of legal person registration; THERESHOLD2 is the risk characteristic value corresponding to the frequency of increased persons, that is, the variable value of the frequency of increased persons; THERESHOLD3 is the risk characteristic value corresponding to the proportion of insured persons in other places, that is, the variable value of the proportion of insured persons in other places.

[0123] Model fitting and evaluation: Use appropriate statistical software (such as R, Python, etc.) to perform regression analysis on the data to obtain the regression coefficients and their significance test results. The effectiveness of the model is evaluated by the following steps:

[0124] R 2Value: A measure of how well the model fits the data.

[0125] P value: To judge the statistical significance of each independent variable, the significance level α=0.05 is usually selected.

[0126] Residual analysis: Check the normality and independence of the model residuals to verify the model assumptions.

[0127] The process of establishing the loss ratio logistic regression model includes:

[0128] Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years and the risk characteristic value corresponding to each risk characteristic dimension;

[0129] Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total loss ratio (total loss ratio of each corporate business), legal person registration time (corresponding risk characteristic value), frequency of adding people (corresponding risk characteristic value), and proportion of non-local insured persons (corresponding risk characteristic value);

[0130] Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of the dimension;

[0131] The dependent variable is the probability of abnormal insurance (1 means abnormal, 0 means normal), and the independent variables are the risk characteristic values ​​corresponding to each risk characteristic dimension and other variables that may affect the loss ratio (such as the insured amount, industry type, etc.). A loss ratio logistic regression model is established.

[0132] The form of the loss ratio logistic regression model is:

[0133] Logit(P)=β 20 +β 21 THERESHOLD1+β 22 THERESHOLD2+β 23

[0134] THERESHOLD3+…

[0135] Among them, P is the probability of abnormal insurance, β 20 is the intercept term; β 21 , β 22 , β 23is the regression coefficient of each variable, indicating the degree of influence of each risk characteristic value on the total claims ratio; THERESHOLD1 is the risk characteristic value corresponding to the legal person registration time, that is, the legal person registration time variable value; THERESHOLD2 is the risk characteristic value corresponding to the increase in the frequency of people, that is, the increase in the frequency of people variable value; THERESHOLD3 is the risk characteristic value corresponding to the proportion of insured persons in other places, that is, the proportion of insured persons in other places variable value.

[0136] Risk probability calculation: Calculate the abnormal insurance probability of each insurance case based on the logistic regression model, and set a probability threshold (such as 0.75) to determine the normality of the insurance.

[0137] Comprehensive assessment: Combine the results of the multivariate linear regression model with the logistic regression model, use thresholds and risk probabilities to comprehensively determine the insurance risk level, and develop corresponding anti-fraud measures.

[0138] By combining the multivariate linear regression model with the logistic regression model, this solution effectively establishes the mathematical relationship between the risk threshold and the loss ratio, providing a powerful tool for identifying abnormal insurance and helping insurance companies reduce fraud risks.

[0139] The process of establishing the loss ratio decision tree model includes:

[0140] Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years and the risk characteristic value corresponding to each risk characteristic dimension;

[0141] Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total loss ratio (total loss ratio of each corporate business), legal person registration time (corresponding risk characteristic value), frequency of adding people (corresponding risk characteristic value), and proportion of non-local insured persons (corresponding risk characteristic value);

[0142] Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of the dimension;

[0143] Taking the risk characteristic value corresponding to each risk characteristic dimension as the input variable and the total loss ratio as the dependent variable, a loss ratio decision tree model is established. According to the relationship between the total loss ratio of each corporate business policy sample data and the total loss ratio threshold, all corporate business policy sample data are classified into high risk and low risk.

[0144] (1) Overview of the loss ratio decision tree model

[0145] For corporate business policy data that exceeds three thresholds, in addition to the multivariate linear regression model, the relationship between risk dimensions (legal person registration duration, frequency of adding people, and proportion of non-local insured persons) and the claim ratio can also be modeled through a decision tree model to further improve the ability to judge the risk of abnormal insurance. Model selection: Select a suitable decision tree algorithm, such as CART (Classification and Regression Trees), for regression analysis.

[0146] (2) Decision tree model construction:

[0147] Use the decision tree algorithm to build a loss ratio prediction model and visualize the relationship between risk threshold and loss ratio. The specific steps are as follows:

[0148] Feature selection: Legal person registration duration, frequency of adding people, and proportion of non-local insured persons are used as feature inputs.

[0149] Label definition: According to the policy claim ratio, set a reasonable threshold (for example, a claim ratio higher than a certain value is marked as high risk, and a claim ratio lower than a certain value is marked as low risk), and classify the policies into high-risk and low-risk categories.

[0150] Decision tree training: The decision tree algorithm is used to train the model for the input policy data and its features. The best split point is determined by indicators such as information gain or Gini coefficient, and the binary tree structure is gradually constructed. The data set is divided into a training set and a test set. Usually 70% of the data is used for training and 30% for testing. The decision tree model is trained using the training set. The input features include THERESHOLD1, THERESHOLD2, THERESHOLD3 and other related features. Model parameters are set, such as the maximum depth, the minimum number of sample splits, etc., to prevent overfitting.

[0151] Model optimization: Through pruning techniques (such as pre-pruning or post-pruning), we can avoid overfitting of decision trees and improve the generalization ability of the model.

[0152] Using all corporate business policy sample data with risk feature values ​​above the risk threshold and corporate business policy classification, the decision tree algorithm is used to train the loss ratio decision tree model. The performance of the model is evaluated using the test set, and the error between the predicted total loss ratio and the actual loss ratio is calculated. Evaluation indicators such as mean square error (MSE), determination coefficient (R 2 ) to evaluate the fit of the model.

[0153] (3) Decision rule generation:

[0154] A decision tree model generates a series of decision rules, such as:

[0155] The first type is to screen out the corporate business insurance policies to be identified whose risk characteristic values ​​corresponding to each risk characteristic dimension are higher than the risk threshold corresponding to each risk characteristic dimension;

[0156] Or, the second method is to screen out any of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified that is higher than the risk threshold corresponding to the risk characteristic dimension;

[0157] Or, the third type is to screen out the corporate business policies to be identified whose corresponding risk characteristic weighted values ​​after weighting each risk characteristic dimension are higher than the corresponding risk weighted threshold after weighting the risk characteristic dimension.

[0158] Or, the fourth type, screening out the corporate business insurance policies to be identified, any two of whose risk characteristic values ​​corresponding to each risk characteristic dimension are higher than the risk threshold corresponding to the risk characteristic dimension;

[0159] For example, if the length of legal person registration is lower than the 25th percentile and the frequency of adding people is higher than the 75th percentile, the loss ratio of the legal person business is high risk (corresponding to the fourth type);

[0160] If the proportion of out-of-town insured persons exceeds the 90th percentile and the length of legal person registration is higher than the average, the policy is low risk (corresponding to Type 4 or Type 1).

[0161] (4) Risk assessment:

[0162] The decision tree model is used to predict the risk level of each insurance policy, and a risk assessment report on each legal person's business is generated by classifying the insurance policies that exceed the threshold. The model output is:

[0163] High risk: The decision tree model determines that the policy is a potential risk policy with a high claim ratio.

[0164] Low risk: The decision tree model determines that the policy has a low claim ratio.

[0165] (5) Model effect evaluation:

[0166] The prediction accuracy of the model is evaluated through methods such as cross-validation, the performance of the decision tree model and the linear regression model in claim ratio prediction and abnormal insurance detection is compared, and the most suitable model is selected for practical application.

[0167] By introducing the decision tree model, this solution effectively establishes the mathematical relationship between risk threshold and loss ratio, thereby helping to identify abnormal insurance risks. This model can not only capture the complex relationship between nonlinear features, but also provide an explainable decision path, providing insurance companies with a more accurate risk management basis.

[0168] Among them, in step S3, the expected loss ratio corresponding to the corporate business policy is determined by using the pre-established loss ratio model and the risk characteristic value corresponding to each risk characteristic dimension:

[0169] Obtaining the risk characteristic value corresponding to each risk characteristic dimension in the insurance policy of the legal person business to be identified, and inputting the risk characteristic value corresponding to each risk characteristic dimension in the insurance policy of the legal person business to be identified into a pre-established loss ratio decision tree model, or a loss ratio logistic regression model, or a loss ratio regression model;

[0170] The total loss ratio output by the pre-established loss ratio decision tree model, or the abnormal insurance probability output by the loss ratio logistic regression model, or the total loss ratio output by the loss ratio regression model is used as the expected loss ratio of the corporate business policy.

[0171] Preferably, a comprehensive evaluation can also be performed based on each loss ratio model. For example, the results of the multivariate linear regression model and the logistic regression model can be combined to comprehensively judge the risk level of the insurance through thresholds and risk probabilities, and formulate corresponding anti-fraud measures.

[0172] By combining the multivariate linear regression model with the logistic regression model, this solution effectively establishes the mathematical relationship between the risk threshold and the loss ratio, providing a powerful tool for identifying abnormal insurance and helping insurance companies reduce fraud risks.

[0173] In step S4, an exception report is output: it contains the contract IDs and detailed information of all contracts whose expected loss ratios are higher than the preset loss ratio threshold, including:

[0174] The registration duration of the policyholder organization.

[0175] Details of the addition or removal of insured persons (including the number and time of addition or removal).

[0176] The regional distribution of the insured persons (such as the proportion of the insured persons in each region).

[0177] Statistics: Provides the number of overall abnormal contracts, statistical indicators for each abnormal dimension, etc.

[0178] Preferably, the data can be grouped according to the time_period parameter. Use the following logic:

[0179] If time_period is "month", the data is grouped by month and the insurance contracts for each month are counted.

[0180] If time_period is "quarter", the data is grouped by quarter and the insurance contracts for each quarter are counted.

[0181] Flag contracts with expected loss ratios above a preset loss ratio threshold as anomalies and alert relevant managers for further investigation.

[0182] Calculate statistical indicators such as the total number of abnormal contracts, the number of abnormal contracts in each dimension, and summarize them in the report.

[0183] Generate an abnormal insurance risk report, listing the contract ID, expected loss ratio, actual loss ratio and corresponding risk assessment results of all abnormal contracts.

[0184] The technical solution of the present invention can effectively save the manual workload of anti-fraud investigation work under the background of frequent organized group insurance fraud, accurately and effectively identify the risk of group accident insurance underwriting fraud, and effectively reduce operating costs.

[0185] In the technical solution of the present invention, the associated risk characteristic dimensions are screened according to the correlation with the claims ratio, and the risk threshold corresponding to each risk characteristic dimension is determined according to the screened risk characteristic dimensions; the corporate business policy to be identified is obtained, and the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified is calculated; according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified and the risk threshold corresponding to each risk characteristic dimension, the corporate business policies with risk characteristic values ​​higher than the risk threshold are screened; for the corporate business policies with risk characteristic values ​​higher than the risk threshold, the expected claims ratio corresponding to the corporate business policy is determined through a pre-established claims ratio model and the risk characteristic value corresponding to each risk characteristic dimension; the corporate business policies with expected claims ratio higher than the preset claims ratio threshold are output, which effectively solves the problem of low timeliness and reliability of fraud detection of group accident insurance caused by the prior art, and effectively improves the timeliness and reliability of fraud detection of group accident insurance.

[0186] The technical solution of the present invention is based on group accident insurance policies in the past three years, and establishes a data set consisting of feature dimensions and claims ratios; calculates the covariance of each feature dimension and the claims ratio, and retains the feature dimensions corresponding to all non-zero covariances as risk feature dimensions; calculates the correlation coefficient between each risk feature dimension and the claims ratio; sorts them according to the absolute value of the correlation coefficient, and screens out several risk feature dimensions with the highest correlation as associated risk feature dimensions, thereby ensuring the reliability of fraud detection of group accident insurance.

[0187] The technical solution of the present invention screens out all the insured persons whose legal person registration time of the insured enterprise is within two years at the time of insurance, and calculates the simple claim ratio of each screened insured person in the past three years; determines the monthly registration value of the insured person whose simple claim ratio exceeds 100%, and establishes a legal person registration time data set, an added person frequency data set, and a non-local insured person proportion data set; based on the quantile method or standard deviation method of the data in the legal person registration time data set, the added person frequency data set, and the non-local insured person proportion data set, the risk threshold corresponding to each characteristic dimension is determined, and traditional statistical methods can be used to perform statistical analysis on limited sample statistical data to determine the risk threshold of the associated risk characteristic dimension, thereby ensuring the reliability of fraud detection in group accident insurance.

[0188] In the technical solution of the present invention, based on the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified and the risk threshold value corresponding to each risk characteristic dimension, screening corporate business policies with risk characteristic values ​​higher than the risk threshold value specifically includes: screening out corporate business policies whose risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified are all higher than the risk threshold value corresponding to each risk characteristic dimension; or screening out corporate business policies whose any one of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified is higher than the risk threshold value corresponding to the risk characteristic dimension; or screening out corporate business policies whose weighted risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified after weighting is higher than the risk weighted threshold value corresponding to the weighted risk characteristic dimension. According to risk requirements, corporate business policies with risk characteristic values ​​higher than the risk threshold value can be flexibly selected, thereby improving the scope of application of fraud detection in group accident insurance.

[0189] In the technical solution of the present invention, the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is obtained, and the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is input into a pre-established loss ratio decision tree model, or a loss ratio logistic regression model, or a loss ratio regression model; the total loss ratio output by the pre-established loss ratio decision tree model, or the abnormal insurance probability output by the loss ratio logistic regression model, or the total loss ratio output by the loss ratio regression model is used as the expected loss ratio of the corporate business policy. The loss ratio model can be flexibly selected according to actual requirements, thereby improving the scope of application and flexibility of fraud detection in group accident insurance.

[0190] Embodiment 2

[0191] like Figure 5 As shown, the technical solution of the present invention also provides a group accident insurance fraud detection system based on abnormal insurance, including:

[0192] The determination module 101 selects the associated risk characteristic dimensions according to the correlation with the loss ratio, and determines the risk threshold corresponding to each risk characteristic dimension according to the selected risk characteristic dimensions; wherein the selected associated risk characteristic dimensions include the registration time of the legal person, the frequency of adding people, and the proportion of non-local insured persons;

[0193] The screening module 102 obtains the corporate business insurance policy to be identified, calculates the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policy to be identified, and screens the corporate business insurance policy whose risk characteristic value is higher than the risk threshold according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policy to be identified and the risk threshold corresponding to each risk characteristic dimension;

[0194] The calculation module 103 determines the expected loss ratio of the corporate business insurance policy for which the risk characteristic value is higher than the risk threshold through the pre-established loss ratio model and the risk characteristic value corresponding to each risk characteristic dimension;

[0195] The output module 104 outputs corporate business policies whose expected loss ratio is higher than a preset loss ratio threshold.

[0196] The implementation process of the determination module 101, the screening module 102, the calculation module 103, and the output module 104 in the second embodiment corresponds to the method steps in the first embodiment, and will not be described in detail in the second embodiment.

[0197] In the technical solution of the present invention, the associated risk characteristic dimensions are screened according to the correlation with the claims ratio, and the risk threshold corresponding to each risk characteristic dimension is determined according to the screened risk characteristic dimensions; the corporate business policy to be identified is obtained, and the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified is calculated; according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified and the risk threshold corresponding to each risk characteristic dimension, the corporate business policies with risk characteristic values ​​higher than the risk threshold are screened; for the corporate business policies with risk characteristic values ​​higher than the risk threshold, the expected claims ratio corresponding to the corporate business policy is determined through a pre-established claims ratio model and the risk characteristic value corresponding to each risk characteristic dimension; the corporate business policies with expected claims ratio higher than the preset claims ratio threshold are output, which effectively solves the problem of low timeliness and reliability of fraud detection of group accident insurance caused by the prior art, and effectively improves the timeliness and reliability of fraud detection of group accident insurance.

[0198] The technical solution of the present invention is based on group accident insurance policies in the past three years, and establishes a data set consisting of feature dimensions and claims ratios; calculates the covariance of each feature dimension and the claims ratio, and retains the feature dimensions corresponding to all non-zero covariances as risk feature dimensions; calculates the correlation coefficient between each risk feature dimension and the claims ratio; sorts them according to the absolute value of the correlation coefficient, and screens out several risk feature dimensions with the highest correlation as associated risk feature dimensions, thereby ensuring the reliability of fraud detection of group accident insurance.

[0199] The technical solution of the present invention screens out all the insured persons whose legal person registration time of the insured enterprise is within two years at the time of insurance, and calculates the simple claim ratio of each screened insured person in the past three years; determines the monthly registration value of the insured person whose simple claim ratio exceeds 100%, and establishes a legal person registration time data set, an added person frequency data set, and a non-local insured person proportion data set; based on the quantile method or standard deviation method of the data in the legal person registration time data set, the added person frequency data set, and the non-local insured person proportion data set, the risk threshold corresponding to each characteristic dimension is determined, and traditional statistical methods can be used to perform statistical analysis on limited sample statistical data to determine the risk threshold of the associated risk characteristic dimension, thereby ensuring the reliability of fraud detection in group accident insurance.

[0200] In the technical solution of the present invention, based on the risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified and the risk threshold value corresponding to each risk characteristic dimension, screening corporate business policies with risk characteristic values ​​higher than the risk threshold value specifically includes: screening out corporate business policies whose risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified are all higher than the risk threshold value corresponding to each risk characteristic dimension; or screening out corporate business policies whose any one of the risk characteristic values ​​corresponding to each risk characteristic dimension of the corporate business policy to be identified is higher than the risk threshold value corresponding to the risk characteristic dimension; or screening out corporate business policies whose weighted risk characteristic value corresponding to each risk characteristic dimension of the corporate business policy to be identified after weighting is higher than the risk weighted threshold value corresponding to the weighted risk characteristic dimension. According to risk requirements, corporate business policies with risk characteristic values ​​higher than the risk threshold value can be flexibly selected, thereby improving the scope of application of fraud detection in group accident insurance.

[0201] In the technical solution of the present invention, the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is obtained, and the risk feature value corresponding to each risk feature dimension in the corporate business policy to be identified is input into a pre-established loss ratio decision tree model, or a loss ratio logistic regression model, or a loss ratio regression model; the total loss ratio output by the pre-established loss ratio decision tree model, or the abnormal insurance probability output by the loss ratio logistic regression model, or the total loss ratio output by the loss ratio regression model is used as the expected loss ratio of the corporate business policy. The loss ratio model can be flexibly selected according to actual requirements, thereby improving the scope of application and flexibility of fraud detection in group accident insurance.

[0202] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A group accident insurance fraud detection method based on abnormal insurance, characterized in that: include: According to the correlation with the loss ratio, the associated risk characteristic dimensions are screened, and the risk threshold corresponding to each risk characteristic dimension is determined based on the screened risk characteristic dimensions; among which, the screened associated risk characteristic dimensions include the length of legal person registration, the frequency of adding people, and the proportion of non-local insured persons; Obtain the corporate business insurance policies to be identified, calculate the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified, and screen corporate business insurance policies whose risk characteristic values ​​are higher than the risk threshold according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified and the risk threshold corresponding to each risk characteristic dimension; For corporate business policies with risk characteristic values ​​higher than the risk threshold, the expected loss ratio of the corporate business policy is determined through the pre-established loss ratio model and the risk characteristic values ​​corresponding to each risk characteristic dimension; Output corporate business policies whose expected loss ratio is higher than the preset loss ratio threshold.

2. According to claim 1, a method for detecting group accident insurance fraud based on abnormal insurance, characterized in that: The dimensions of screening associated risk characteristics based on the correlation with the loss ratio specifically include: Based on the group accident insurance policies in the past three years, a data set consisting of feature dimensions and claim ratios was established; Calculate the covariance between each feature dimension and the loss ratio, and retain all feature dimensions corresponding to non-zero covariances as risk feature dimensions; Calculate the correlation coefficient between each risk characteristic dimension and the loss ratio; wherein the correlation coefficient is the ratio of the first calculation factor to the second calculation factor, the first calculation factor is the covariance of each risk characteristic dimension and the loss ratio, and the second calculation factor is the standard deviation of each risk characteristic dimension and the loss ratio; Sort by the absolute value of the correlation coefficient and select several risk feature dimensions with the highest correlation as associated risk feature dimensions.

3. A group accident insurance fraud detection method based on abnormal insurance according to claim 1, characterized in that: Determining the risk threshold corresponding to each risk feature dimension according to the screening risk feature dimension includes: Screen out all the insured persons whose legal person registration time of the insured enterprise is within two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years; among them, the insured person of group accident insurance is the legal person insured; Determine the monthly registration value of the insured with a simple claim ratio exceeding 100%, and establish a legal person registration duration data set; determine the risk threshold corresponding to the legal person registration duration based on the quantile method or standard deviation method of the legal person registration duration data set; Screen out the insurance policies of the insured with simple claims ratio exceeding 100%, calculate the ratio of the number of people added or reduced in the insurance policies after screening to the total number of insured persons during the effective period of the insurance policies, and establish a frequency data set of people added; determine the risk threshold corresponding to the frequency of people added based on the quantile method or standard deviation method of the frequency data of people added in the frequency data set of people added; Calculate the ratio of all non-local insured persons to the total number of insured persons during the validity period of the screened policies, and establish a non-local insured person ratio data set; determine the risk threshold corresponding to the non-local insured person ratio based on the quantile method or standard deviation method of the non-local insured person ratio data set; where non-local insured persons are insured persons whose registered address is not in the same prefecture-level city as the underwriting enterprise of the group accident insurance.

4. A group accident insurance fraud detection method based on abnormal insurance according to claim 3, characterized in that: Based on the quantile method or standard deviation method of the legal person registration time data set, the risk threshold corresponding to the legal person registration time is determined as follows: The risk threshold corresponding to the legal person registration time is: the 25th percentile of the legal person registration time data in the legal person registration time data set, or the difference between the mean of the legal person registration time data in the legal person registration time data set and the geometric multiple standard deviation; Based on the quantile method or standard deviation method of the increased frequency data of the increased frequency data set, the risk threshold corresponding to the increased frequency is determined as follows: The risk threshold corresponding to the increased person frequency is: the 75th percentile of the increased person frequency data of the increased person frequency data set, or the sum of the mean and geometric multiple standard deviation of the legal person registration duration data of the legal person registration duration data set; Based on the quantile method or standard deviation method of the non-local insured proportion data set, the risk threshold corresponding to the non-local insured proportion is determined as follows: The risk threshold corresponding to the proportion of non-local insured persons is: the 90th percentile of the non-local insured persons proportion data of the non-local insured persons proportion data set, or the sum of the mean of the non-local insured persons proportion data of the non-local insured persons proportion data set and the geometric multiple standard deviation.

5. The group accident insurance fraud detection method based on abnormal insurance according to claim 1 is characterized in that: Obtain the insurance policy of the legal person business to be identified, and calculate the risk characteristic value corresponding to each risk characteristic dimension of the insurance policy of the legal person business to be identified, including: Obtain the insured information dataset in the insurance policy of the legal person to be identified. The insured information includes the insurance date, the registration date of the insured legal person, the date of adding or removing the insured, and the insured ID number; Calculate the risk characteristic value corresponding to the legal person registration time based on the insurance date and the insured legal person registration date; Calculate the risk characteristic value corresponding to the frequency of adding people after the insurance date but before the insurance takes effect based on the insurance date and the date of adding or reducing the insured person; The risk characteristic value corresponding to the proportion of non-local insured persons is calculated based on the insured’s ID number.

6. A group accident insurance fraud detection method based on abnormal insurance according to claim 1, characterized in that: According to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policy to be identified and the risk threshold corresponding to each risk characteristic dimension, the corporate business insurance policies with risk characteristic values ​​higher than the risk threshold are screened, including: Screen out the corporate business insurance policies to be identified whose risk characteristic values ​​corresponding to each risk characteristic dimension are higher than the risk threshold corresponding to each risk characteristic dimension; Or, screen out the corporate business insurance policies to be identified whose risk characteristic values ​​corresponding to each risk characteristic dimension are higher than the risk threshold corresponding to the risk characteristic dimension; Alternatively, screen out the corporate business policies to be identified whose corresponding risk characteristic weighted values ​​after weighting each risk characteristic dimension are higher than the corresponding risk weighted threshold after weighting the risk characteristic dimension.

7. According to the group accident insurance fraud detection method based on abnormal insurance according to claim 1, the pre-established loss ratio model includes a loss ratio regression model, a loss ratio logistic regression model, and a loss ratio decision tree model.

8. A method for detecting group accident insurance fraud based on abnormal insurance according to claim 7, characterized in that: The process of establishing the loss ratio regression model includes: Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years and the risk characteristic value corresponding to each risk characteristic dimension; Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total claim ratio, corporate registration time, frequency of adding people, and proportion of non-local insured persons; Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of the dimension; Taking the total loss ratio as the dependent variable and the risk characteristic value corresponding to each risk characteristic dimension as the independent variable, a multivariate linear regression model of the loss ratio is established; The process of establishing the loss ratio logistic regression model includes: Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years and the risk characteristic value corresponding to each risk characteristic dimension; Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total claim ratio, corporate registration time, frequency of adding people, and proportion of non-local insured persons; Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of the dimension; The dependent variable is the probability of abnormal insurance, and the independent variables are the risk characteristic values ​​corresponding to each risk characteristic dimension and other variables that may affect the loss ratio, and a loss ratio logistic regression model is established; The process of establishing the loss ratio decision tree model includes: Screen out all the insured persons whose legal person registration time of the insured enterprise is less than two years at the time of insurance, and calculate the simple claim ratio of each screened insured person in the past three years and the risk characteristic value corresponding to each risk characteristic dimension; Screen all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold. Each corporate business policy sample data includes: total claim ratio, corporate registration time, frequency of adding people, and proportion of non-local insured persons; Standardize all corporate business policy sample data with risk characteristic values ​​higher than the risk threshold to eliminate the impact of the dimension; Taking the risk characteristic value corresponding to each risk characteristic dimension as the input variable and the total loss ratio as the dependent variable, a loss ratio decision tree model is established. According to the relationship between the total loss ratio of each corporate business policy sample data and the total loss ratio threshold, all corporate business policy sample data are classified into high risk and low risk. Using the sample data of all corporate business insurance policies with risk feature values ​​higher than the risk threshold and the classification of corporate business insurance policies, the decision tree algorithm is used to train the loss ratio decision tree model.

9. The group accident insurance fraud detection method based on abnormal insurance according to claim 7 is characterized in that: The expected loss ratio of the insurance policy for the legal person business is determined by the pre-established loss ratio model and the risk characteristic value corresponding to each risk characteristic dimension: Obtaining the risk characteristic value corresponding to each risk characteristic dimension in the insurance policy of the legal person business to be identified, and inputting the risk characteristic value corresponding to each risk characteristic dimension in the insurance policy of the legal person business to be identified into a pre-established loss ratio decision tree model, or a loss ratio logistic regression model, or a loss ratio regression model; The total loss ratio output by the pre-established loss ratio decision tree model, or the abnormal insurance probability output by the loss ratio logistic regression model, or the total loss ratio output by the loss ratio regression model is used as the expected loss ratio of the corporate business policy.

10. A group accident insurance fraud detection system based on abnormal insurance, characterized in that: include: The determination module selects the associated risk characteristic dimensions according to the correlation with the loss ratio, and determines the risk threshold corresponding to each risk characteristic dimension according to the selected risk characteristic dimensions; wherein the associated risk characteristic dimensions after screening include the registration time of the legal person, the frequency of adding people, and the proportion of non-local insured persons; The screening module obtains the corporate business insurance policies to be identified, calculates the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified, and screens the corporate business insurance policies whose risk characteristic values ​​are higher than the risk threshold according to the risk characteristic value corresponding to each risk characteristic dimension of the corporate business insurance policies to be identified and the risk threshold corresponding to each risk characteristic dimension; The calculation module determines the expected loss ratio of the corporate business insurance policy for which the risk characteristic value is higher than the risk threshold through the pre-established loss ratio model and the risk characteristic value corresponding to each risk characteristic dimension; The output module outputs corporate business policies whose expected claims ratio is higher than the preset claims ratio threshold.

Images