Personal insurance customer fraud risk assessment method and device based on machine learning correction

Abstract

The invention provides a personal insurance customer fraud risk assessment method and device based on machine learning correction, and the method comprises the following steps: obtaining risk information of a customer, the risk information comprises insurance type information, personal attribute information and insurance change information, calculating a product risk score based on the insurance type information, calculating an initial personal attribute risk score based on the personal attribute information, and calculating a dynamic behavior score based on the insurance change information; adjusting the initial personal attribute risk score to obtain a corrected personal attribute risk score; and taking the sum of the product risk score and the corrected personal attribute risk score as a basic risk score, and performing weighted summation on the basic risk score and the dynamic behavior score to obtain a risk score of the customer. According to the scheme, the initial personal attribute risk score of the score card model is corrected through the feature weight of the risk assessment model, the dependence on artificial experience is weakened, the risk weight is matched with an actual data rule, and the assessment accuracy is improved while the risk judgment basis is visually displayed.

Description

Technical Field

[0001] This application relates to the field of risk management technology, and in particular to a method and apparatus for assessing fraud risk of life insurance customers based on machine learning correction. Background Technology

[0002] Against the backdrop of the rapid development of the life insurance industry, the customer base is becoming increasingly diverse, and business scenarios are constantly innovating. Accurate assessment and dynamic management of customer money laundering risks have become a core requirement for industry compliance management. On the one hand, life insurance business involves multi-dimensional compliance requirements such as customer identity verification, fund flow monitoring, and business conduct standardization. Insurance institutions need to use effective technological means to identify potential money laundering risks from customers and ensure compliant business operations. On the other hand, with the expansion of business scale and the increasing complexity of customer behavior, traditional risk assessment methods are gradually revealing their inadequacy and failing to meet the needs of refined compliance management.

[0003] Current mainstream customer money laundering risk assessment solutions in the industry have significant limitations. Firstly, most institutions rely on fixed rule engines built upon regulatory requirements for risk identification. These engines use pre-set customer identity verification standards, thresholds for specific transaction amounts, and basic abnormal behavior screening conditions to conduct risk checks on customer policy information. However, the assessment rules of these solutions are relatively rigid, covering a limited range of risk characteristics. They cannot adapt promptly to new risk scenarios arising from new product forms and business models, and lack the ability to proactively adjust and optimize based on changing risk trends, easily leading to delayed or missed risk identification.

[0004] Secondly, some institutions use customer risk scoring card models for risk quantification. These models analyze key indicators such as customer attributes, product types, and transaction records, assigning different weights based on expert experience to calculate customer risk scores. However, the risk feature selection and weighting of these models heavily rely on human experience, making them subjective and unable to objectively reflect the actual impact of risk factors in different business scenarios. For example, when assessing dimensions such as the matching degree between customer premium expenditure and income, and the risk associated with occupational type, the experience-based weighting may deviate from the actual risk distribution, resulting in insufficient accuracy in risk assessment and an inability to effectively distinguish between low-risk and medium-to-high-risk customers.

[0005] In recent years, while some organizations have attempted to introduce machine learning algorithms to optimize risk assessment by training models with historical customer data and uncovering potential risk correlations within the data, the risk assessment process of these purely algorithmic models is inherently "black box"-like. The core logic of the model's decisions, such as the mechanisms by which key risk characteristics function and the basis for calculating risk probabilities, is difficult to clearly present and trace. Furthermore, the compliance management field has extremely high requirements for the interpretability and verifiability of risk assessment results, necessitating a clear explanation of the basis for risk determination to regulatory agencies and internal management departments. This limits the practical application of purely algorithmic models in real-world compliance scenarios.

[0006] In summary, the life insurance industry urgently needs to overcome the bottlenecks of "rigid rules, reliance on experience, and insufficient interpretability" in the field of customer money laundering risk assessment. It needs to build a data-driven, logically traceable, and dynamically adaptable assessment system to improve the accuracy and flexibility of money laundering risk identification and meet the dual needs of deepening industry compliance management and high-quality business development. Summary of the Invention

[0007] This application provides a method and apparatus for assessing personal insurance customer fraud risk based on machine learning correction. It corrects the initial personal attribute risk score of the scoring card model by adjusting the feature weights of the risk assessment model, weakens the reliance on human experience, makes the risk weights match the actual data patterns, and improves the accuracy of the assessment while intuitively displaying the basis for risk judgment.

[0008] In a first aspect, embodiments of this application provide a method for assessing fraud risk of life insurance customers based on machine learning correction, the method comprising:

[0009] Obtain customer risk information, including insurance type information, personal attribute information, and insurance change information; calculate product risk score based on insurance type information; pre-train a scoring card model; input personal attribute information into the scoring card model; the scoring card model calculates an initial personal attribute risk score based on preset scoring card weights; and calculate a dynamic behavior score based on insurance change information. The model feature weights corresponding to personal attribute information in the pre-trained risk assessment model are extracted. The ratio of the model feature weights to the scorecard weights is used as the adjustment coefficient. The initial personal attribute risk score is adjusted using the adjustment coefficient to obtain the corrected personal attribute risk score. A binary classification architecture is constructed. The binary classification architecture is set to predict whether a customer is a money laundering risk user or a non-money laundering risk user based on the customer's risk information. The risk information of multiple historical customers is used as the training dataset. The binary classification architecture is trained using the training dataset to obtain the trained risk assessment model. The base risk score is obtained by weighting and summing the product risk score and the modified personal attribute risk score.

[0010] Secondly, embodiments of this application provide a personal insurance customer fraud risk assessment device based on machine learning correction, comprising: The acquisition module is used to acquire the customer's risk information, which includes insurance type information, personal attribute information and insurance change information. Based on the insurance type information, the product risk score is calculated, a pre-trained scoring card model is performed, and personal attribute information is input into the scoring card model. The scoring card model calculates an initial personal attribute risk score based on the personal attribute information based on the preset scoring card weights, and calculates a dynamic behavior score based on the insurance change information. The correction module extracts the model feature weights corresponding to personal attribute information from the pre-trained risk assessment model, uses the ratio of model feature weights to scorecard weights as adjustment coefficients, and adjusts the initial personal attribute risk score using the adjustment coefficients to obtain the corrected personal attribute risk score. In this module, a binary classification architecture is constructed, which is set to predict whether a customer is a money laundering risk user or a non-money laundering risk user based on the customer's risk information. The risk information of multiple historical customers is used as a training dataset, and the binary classification architecture is trained using the training dataset to obtain the trained risk assessment model. The assessment module uses the sum of the product risk score and the modified personal attribute risk score as the base risk score. The base risk score is then weighted and summed with the dynamic behavior score to obtain the customer's risk score.

[0011] Thirdly, embodiments of this application provide an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform a method for assessing personal insurance customer fraud risk based on machine learning correction.

[0012] Fourthly, embodiments of this application provide a readable storage medium storing a computer program that, when executed by a processor, implements a method for assessing personal insurance customer fraud risk based on machine learning correction.

[0013] The main contributions and innovations of this invention are as follows: This solution calculates multi-dimensional risk scores by integrating insurance product information, personal attribute information, and insurance change information, comprehensively covering inherent product risks, inherent customer attribute risks, and dynamic behavioral risks, avoiding the limitations of single-dimensional assessment. The solution corrects the initial personal attribute risk scores of the scoring card model through feature weights in the risk assessment model, reducing reliance on human experience and allowing risk weights to match actual data patterns, intuitively displaying the basis for risk judgment while improving assessment accuracy. This solution automatically calculates dynamic behavioral scores daily based on insurance change information, capturing changes in customer behavior in real time, adapting promptly to new risk scenarios, and solving the problem of rigid traditional rules.

[0014] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a machine learning-based method for assessing fraud risk in life insurance customers, according to an embodiment of this application. Figure 2 This is a structural block diagram of a personal insurance customer fraud risk assessment device based on machine learning correction, according to an embodiment of this application. Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0017] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0018] Example 1 This application provides a method for assessing fraud risk in personal insurance customers based on machine learning. It corrects the initial personal attribute risk score of the scoring card model by adjusting the feature weights of the risk assessment model, reducing reliance on human experience, allowing risk weights to match actual data patterns, and intuitively displaying the basis for risk judgment while improving assessment accuracy. Specifically, refer to... Figure 1 The method includes: Obtain customer risk information, including insurance type information, personal attribute information, and insurance change information; calculate product risk score based on insurance type information; pre-train a scoring card model; input personal attribute information into the scoring card model; the scoring card model calculates an initial personal attribute risk score based on preset scoring card weights; and calculate a dynamic behavior score based on insurance change information. The model feature weights corresponding to personal attribute information in the pre-trained risk assessment model are extracted. The ratio of the model feature weights to the scorecard weights is used as the adjustment coefficient. The initial personal attribute risk score is adjusted using the adjustment coefficient to obtain the corrected personal attribute risk score. A binary classification architecture is constructed. The binary classification architecture is set to predict whether a customer is a money laundering risk user or a non-money laundering risk user based on the customer's risk information. The risk information of multiple historical customers is used as the training dataset. The binary classification architecture is trained using the training dataset to obtain the trained risk assessment model. The base risk score is obtained by weighting and summing the product risk score and the modified personal attribute risk score, and then summing the base risk score and the dynamic behavior score.

[0019] In this plan, "customer" specifically refers to the policyholder or prospective policyholder of a life insurance product.

[0020] In the current embodiment, the insurance information is the basic information of the insurance product that the customer wants to purchase. The insurance information corresponds to at least one insurance assessment dimension. The customer's insurance score in each insurance assessment dimension is obtained based on the preset insurance scoring rules, and the sum of the insurance scores of all insurance assessment dimensions is used as the customer's product risk score.

[0021] Specifically, the insurance information mentioned in this plan includes four insurance assessment dimensions: product type, payment method, premium amount, and source of funds. The scoring rules for each insurance assessment dimension are defined in the preset scoring rules.

[0022] For example, product types include high-risk, medium-risk, and low-risk products. High-risk products are characterized by medium or high complexity, with no guarantee of fluctuating policy benefits, or high complexity, with guaranteed fluctuations in policy benefits; corresponding products are investment-linked life insurance, variable annuity insurance, and other life insurance products that meet this description, including whole life insurance, annuity insurance, and endowment insurance. Medium-risk products are characterized by medium complexity, with guaranteed fluctuations in policy benefits; corresponding products are dedicated commercial pension insurance, participating life insurance, universal life insurance, and other life insurance products that meet this description, including life insurance and annuity insurance. Low-risk products are characterized by low or medium complexity, with certain policy benefits, including life insurance, annuity insurance, health insurance, and accident insurance.

[0023] In the preset scoring rules, high-risk products are defined as scoring 15 points, medium-risk products as scoring 10 points, and low-risk products as scoring 5 points.

[0024] Payment methods include lump sum payment and installment payment. In the preset scoring rules, lump sum payment is defined as 15 points and installment payment as 5 points.

[0025] The premium amount includes amounts of RMB 1 million or more, RMB 500,000 or more, and RMB 100,000 or more. In the preset scoring rules, RMB 1 million or more is defined as scoring 20 points, RMB 500,000 or more as scoring 15 points, and RMB 100,000 or more as scoring 5 points.

[0026] The sources of funds include cash payments, third-party payments from non-immediate family members, and bank transfers from the account of the individual or their immediate family members. In the preset scoring rules, cash payments are defined as scoring 20 points, third-party payments from non-immediate family members as scoring 15 points, and bank transfers from the account of the individual or their immediate family members as scoring 5 points.

[0027] For example, if a customer's insurance information is as follows: the product type is investment-linked insurance, the payment method is lump sum, the premium amount is RMB 1.2 million (greater than or equal to RMB 1 million), and the source of funds is a third-party payment from a non-immediate family member, then the sum of the insurance scores across all insurance assessment dimensions will be used as the customer's product risk score, i.e., 15+15+20+15=65 points.

[0028] In the current embodiment, the personal attribute information corresponds to at least one personal attribute assessment dimension. The scoring card model maps each personal attribute assessment dimension to a corresponding personal attribute score based on a preset scoring card weight, and uses the sum of the personal attribute scores of all personal attribute assessment dimensions as the customer's initial personal attribute risk score.

[0029] Specifically, since personal attribute information is messy and the importance of different personal attribute information cannot be determined, it is impossible to directly formulate corresponding scoring rules for scoring. The scorecard model is a tool to quantify complex risk factors into specific scores. Therefore, this solution uses the risk information of multiple historical customers as a training dataset to pre-train the scorecard model to obtain scorecard weights, and then uses the score to quantify the risk level of customer personal attribute information based on the mapping of scorecard weights.

[0030] Specifically, personal attribute information includes five personal attribute assessment dimensions: age, occupation, income-premium matching degree, identity, and regional risk. During the training process, the scorecard model learns the importance of different personal attribute assessment dimensions to customer risk and generates scorecard weights. By formulating mapping rules, different personal attribute assessment dimensions are mapped to different personal attribute scores.

[0031] For example, the age range includes men over 65 years old, women over 60 years old, women under 25 years old, and other age ranges. In the mapping rules, men over 65 years old and women over 60 years old score 15 points, women under 25 years old score 10 points, and other age ranges score 0 points. Specifically, men over 65 years old and women over 60 years old are more concerned about asset transfer and may involve complex and unconventional financial arrangements, so their personal attribute scores are high; women under 25 years old have a short income accumulation period, which does not match the high premium expenditure, and the source of funds needs to be carefully verified, so their personal attribute scores are medium; other age ranges are the conventional risk range, so their personal attribute scores are low.

[0032] Occupations include those without stable employment (unemployed or without a fixed employer), cash-intensive industries (entertainment, jewelry, or individual retail), high-risk industries, and stable occupations (civil servants, teachers, doctors, etc., or corporate employees, etc.). In the mapping rules, those without stable employment score 15 points, cash-intensive industries score 10 points, high-risk industries score 10 points, and stable occupations score 0 points. Specifically, those without stable employment have difficulty verifying their ability to consistently pay high premiums, raising questions about the legality of their funds, hence their high personal attribute score; cash-intensive industries involve large cash transaction volumes, making them more susceptible to illicit funds, hence their moderate personal attribute score; high-risk industries inherently carry higher risks, hence their moderate personal attribute score; and stable occupations offer stability, transparent income sources, and low inherent risk, hence their low personal attribute score.

[0033] The income-premium matching score includes annual income < annual premium, annual income < twice the annual premium, and annual income ≥ three times the annual premium. In the mapping rules, annual income < annual premium scores 25 points, annual income < twice the annual premium scores 15 points, and annual income ≥ three times the annual premium scores 0 points. Specifically, income < annual premium indicates a severe mismatch between income and premium, meaning the premium funds do not originate from the claimed normal income, hence a high personal attribute score; annual income < twice the annual premium indicates a significant gap between payment ability and premium expenditure, requiring a reasonable explanation of the source of funds, hence a moderate personal attribute score; annual income ≥ three times the annual premium indicates that premium expenditure is within a reasonable range of affordability and low risk, hence a low personal attribute score.

[0034] Identity categories include key positions, non-residents, corporate or entity shareholders, and ordinary residents. In the mapping rules, key positions score 20 points, non-residents score 10 points, corporate or entity shareholders score 10 points, and ordinary residents score 0 points. Specifically, key positions have difficulty verifying their funding sources, posing a significant risk, hence their high personal attribute score; non-residents may have cross-border flows, resulting in a moderate personal attribute score; corporate or entity shareholders may conceal fund flows through complex corporate structures, also resulting in a moderate personal attribute score; and ordinary residents have simple and transparent identities, representing lower risk, hence their low personal attribute score.

[0035] Geographic risk includes high-risk countries, high-risk regions, virtual addresses, and normal addresses. In the mapping rules, high-risk countries score 15 points, high-risk regions score 10 points, virtual addresses score 5 points, and normal addresses score 0 points. Specifically, high-risk countries have extremely high funding source risks, so their personal attribute scores are high; high-risk regions are areas with a lot of financial activity, so their personal attribute scores are moderate; virtual addresses are difficult to effectively identify, so their personal attribute scores are moderate; and normal addresses indicate that the risk is within a controllable range, so their scores are low.

[0036] For example, the calculation of the initial personal attribute risk score in this scheme is illustrated through different scenarios: Scenario 1: High-Risk Customer Profile Client Background: A 70-year-old male retiree (age +15), without a stable job (occupation +15), claims an annual income of 80,000 yuan, yet purchased an insurance policy with an annual premium of 200,000 yuan (income-premium matching +25). He is the younger brother of a mayor in a certain country (client status +20).

[0037] The client's personal risk profile is then calculated as follows: 15 (age) + 15 (occupation) + 25 (income-premium matching) + 20 (identity) + 0 (regional risk) = 75 points.

[0038] In Scenario 1, the client exhibits extremely high risk across multiple inherent risk dimensions, resulting in a very high client attribute risk score. Therefore, the client should be identified as a high-risk client during the "cold start" phase, and enhanced due diligence should be initiated immediately.

[0039] Scenario 2: Profile of Medium-Risk Customers Client Background: A 28-year-old bar owner (occupation +10), annual income of 300,000 yuan, purchasing insurance with an annual premium of 180,000 yuan (income matching +15), no regional anomalies. The risk score for this customer's personal attributes is 0 (age) + 10 (occupation) + 15 (income matching) + 0 (identity) + 0 (region) = 25 points.

[0040] In scenario two, the client received a score due to a slight mismatch between their occupation and income, placing them in the medium-risk range. Therefore, a more in-depth inquiry and verification of their source of funds is needed when applying for insurance.

[0041] Scenario 3: Low-risk customer profile Client background: A 40-year-old public hospital doctor (occupation 0), with an annual income of 500,000 yuan, purchases health insurance with an annual premium of 50,000 yuan (income matching 0).

[0042] The risk score for this customer's personal attributes is 0 (age) + 0 (occupation) + 0 (income matching) + 0 (identity) + 0 (region) = 0 points.

[0043] In scenario three, the client has a clean background, a stable job, reasonable premium expenditures, and extremely low inherent risk, making a simplified due diligence process applicable.

[0044] In the current embodiment, the insurance change information is the dynamic behavior information of the customer in insurance business. The insurance change information corresponds to at least one insurance change dimension. Based on the preset insurance change scoring rules, the insurance change score of the customer's insurance type information in each insurance change assessment dimension is obtained, and the sum of the insurance change scores of all insurance change assessment dimensions is used as the customer's dynamic behavior.

[0045] Furthermore, the customer's insurance change information is automatically calculated daily.

[0046] Specifically, the insurance change information includes three dimensions: frequency of risk behavior, amount of operation, and changes in risk behavior. The pre-defined insurance change rules define the scoring rules for each insurance change assessment dimension.

[0047] For example, the frequency of risky behaviors is the total number of risk protection behaviors within a set time period. These risk protection behaviors include changing the policyholder / insured / beneficiary, changing the address / contact number / email, receiving survival benefits, taking out a policy loan, surrendering the policy, and increasing / decreasing the coverage. The insurance change rules map the total number of risk protection behaviors to the insurance change score. Specifically, the set time period is the past 3 months. If there are 1-2 risk protection behaviors in the past 3 months, the insurance change score is 5 points; if there are 3-5 risk protection behaviors in the past 3 months, the insurance change score is 15 points; and if there are more than 5 risk protection behaviors in the past 3 months, the insurance change score is 25 points.

[0048] The operation amount is the premium change amount within a set time period. The premium change amount includes the cumulative surrender amount, cumulative policy loan amount, cumulative survival benefit amount, and the premium change amount involved in the cumulative increase / decrease in coverage. In the insurance change rules, the premium change amount is mapped to an absolute amount score and a relative amount score, and the higher score between the absolute amount score and the relative amount score is taken as the corresponding insurance change score. The absolute amount score is obtained based on the size of the premium change amount, and the relative amount score is obtained based on the ratio of the premium change amount to the total premium. Specifically, the time period is set to the past 3 months. In terms of absolute amount score, if the premium change amount in the past 3 months is ≥ 500,000 yuan, it indicates a huge capital flow and a high risk, with an absolute amount score of 25 points; if the premium change amount in the past 3 months is ≥ 200,000 yuan, it indicates a large capital flow that requires close attention, with an absolute amount score of 15 points; if the premium change amount in the past 3 months is ≥ 50,000 yuan, it indicates a significant capital flow, with an absolute amount score of 10 points; if the premium change amount in the past 3 months is < 50,000 yuan, it is a normal transaction amount, with an absolute amount score of 0 points. In the relative amount score, if the ratio of the change in premium amount to the total premium in the past 3 months is ≥80%, it indicates that almost all funds have been realized, and the relative amount score is 25 points; if the ratio is ≥50%, it indicates that more than half of the funds have been utilized, and the relative amount score is 15 points; if the ratio is ≥20%, it indicates that the amount of operation has involved a portion of the policy value, and the relative amount score is 10 points; if the ratio is <20%, it indicates that the amount of operation is normal cash flow, and the relative amount score is 0 points.

[0049] For example, if a customer surrenders their policy within the past three months and receives 500,000 yuan, and their total premium is 600,000 yuan, then their absolute amount score is 15 points, their relative amount score is 25 points, and their insurance change score in the dimension of the transaction amount is 25 points.

[0050] Changes in risk behavior are defined as the growth rate of risk behavior frequency. This growth rate is calculated based on the increase in the frequency of risk behavior in the current set time period compared to the previous set time period. In the insurance change rules, this growth rate is mapped to an insurance change score. Specifically, the set time period is the past 3 months, and the formula for calculating the growth rate of risk behavior frequency is: Risk behavior frequency growth rate = (current 3-month risk behavior frequency - previous 3-month risk behavior frequency) / (previous 3-month risk behavior frequency + 1) 100% If the frequency growth rate of risky behaviors is ≥200%, it indicates a drastic change in the behavior pattern, and the insurance change score is 25 points; if the frequency growth rate of risky behaviors is ≥100%, it indicates a significantly active behavior pattern, and the insurance change score is 15 points; if the frequency growth rate of risky behaviors is ≥50%, it indicates an abnormal behavior pattern, and the insurance change score is 5 points; if the frequency growth rate of risky behaviors is <50%, it indicates a stable behavior pattern, and the insurance change score is 0 points.

[0051] For example, suppose the system performs calculations for customer A on day T+1, summarizing data from the past 3 months (the current window): Insurance change score corresponding to the frequency of risk behavior: a total of 4 risk behaviors (2 policy loans, 1 policy surrender, 1 change of beneficiary) → according to the frequency rule (4 times = 15 points). Insurance change score corresponding to the operation amount: Total policy loan amount of 200,000 yuan (absolute amount ≥ 200,000 yuan: 15 points; relative proportion ≥ 20%: 10 points) → 15 points; Surrender amount of 300,000 yuan, total premium of 750,000 yuan (absolute amount: 15 points; relative proportion 40%: 10 points) → 15 points; Total operation amount score = 15 + 15 = 30 points; Insurance change scores corresponding to behavioral trends: There was only one risky behavior in the previous 3-month window. The growth rate of the number of risky behaviors = (4-1) / (1+1) 100% = 150% → 15 points (significant increase).

[0052] Therefore, Customer A's dynamic behavior score = 15 points + 30 points + 15 points = 60 points.

[0053] Conclusion: Customer A has exhibited frequent and substantial recent activity, with a significant upward trend in activity levels, indicating a very high risk level in their dynamic behavior score. This high score will significantly increase their overall risk rating, potentially classifying them as "high-risk" and triggering corresponding early warning and control measures.

[0054] In the current implementation, the risk assessment model adopts the LightGBM architecture or XGBoost architecture. During the training process of the risk assessment model, users at risk of money laundering are marked as positive samples and users at risk of non-money laundering are marked as negative samples in the labeled training dataset. After training, the model feature weights for each attribute assessment dimension are extracted based on the corresponding personal attribute information.

[0055] Specifically, in the process of labeling training samples, 1 represents users at risk of money laundering and 0 represents users at risk of not laundering money. To solve the problem of sample imbalance, oversampling is used to adjust the model weights. The labels of users at risk of money laundering and users at risk of not laundering money are determined based on the customer's historical information (customers who have been confirmed to have committed money laundering fraud in the past are users at risk of money laundering, and the rest are users at risk of not laundering money).

[0056] Specifically, in addition to using customer risk information for training, derived features are constructed during the training process to ensure the accuracy of obtaining model feature weights. Derivative features include the cross features of age and premium, different product purchase combinations, and statistics of historical behavioral features (such as the number of policy maintenances in the past year).

[0057] Specifically, when using the LightGBM architecture as the risk assessment model, gradient one-sided sampling (GOSS) and mutually exclusive feature binding (EFB) techniques are used for training. Gradient one-sided sampling (GOSS) and mutually exclusive feature binding (EFB) techniques have fast training speed and low memory consumption, making them very suitable for processing massive financial data with high-dimensional features. The training steps when using the LightGBM architecture as the risk assessment model are: loading data -> setting a binary classification objective function -> using GOSS and EFB for efficient training -> tuning parameters through cross-validation.

[0058] Specifically, when using the XGBoost architecture as a risk assessment model, its precise greedy algorithm and second-order Taylor expansion are used to obtain high-precision prediction results, but the training resources are consumed more. The training steps when using the XGBoost architecture as a risk assessment model are: loading data -> setting objective:binary:logistic -> tuning parameters through cross-validation.

[0059] In the current embodiment, the model feature weights of each individual attribute assessment dimension corresponding to the personal attribute information are obtained from the pre-trained risk assessment model, and the scorecard weights of each individual attribute assessment dimension corresponding to the personal attribute information are obtained from the scorecard model. The ratio of the model feature weights to the scorecard weights under the corresponding individual attribute assessment dimension is used as the adjustment coefficient of the corresponding individual attribute assessment dimension. The corresponding individual attribute score is adjusted by the adjustment coefficient of each individual attribute assessment dimension to obtain the adjusted individual attribute score. The sum of the adjusted individual attribute scores of all individual attribute assessment dimensions is used as the customer's corrected individual attribute risk score.

[0060] Specifically, the feature weight parameters of each individual attribute assessment dimension obtained in the risk assessment model are normalized to obtain the model feature weight of each individual attribute assessment dimension, and a mapping relationship is constructed between the scorecard weight of the individual attribute assessment dimension and the model feature weight of the individual attribute assessment dimension to achieve a one-to-one correspondence.

[0061] Specifically, in the step of obtaining the scorecard weights for each individual attribute assessment dimension corresponding to the individual attribute information, the sum of the maximum individual attribute scores for each individual attribute assessment dimension is obtained as the total individual attribute score, and the ratio of the maximum individual attribute score for each individual attribute assessment dimension to the total individual attribute score is used as the scorecard weight for the corresponding individual attribute assessment dimension.

[0062] For example, the normalized model feature weights are represented as follows: age model feature weight: 15%; occupation model feature weight: 25%; income-premium matching model feature weight: 35%; identity model feature weight: 15%; regional risk model feature weight: 10%.

[0063] For example, the total score for the five personal attribute assessment dimensions—age, occupation, income-premium matching, identity, and regional risk—is calculated as 15 (age) + 15 (occupation) + 25 (income-premium matching) + 20 (identity) + 15 (regional risk) = 90 points. The weights of the scorecards are as follows: Age scorecard weight: 15 / 90 ≈ 16.7%; Occupation scorecard weight: 15 / 90 ≈ 16.7%; Income-premium matching scorecard weight: 25 / 90 ≈ 27.8%; Identity scorecard weight: 20 / 90 ≈ 22.2%; Regional risk scorecard weight: 15 / 90 ≈ 16.7%.

[0064] The adjustment coefficient is then expressed as: Adjustment factor for the age personal attribute assessment dimension: 15% / 16.7%≈89.8%; Adjustment factor for the dimensions of professional personal attribute assessment: 25% / 16.7%≈149.7%; Adjustment coefficient for the personal attribute evaluation dimension of premium income matching: 35% / 27.8% ≈ 125.9%; Adjustment coefficient for the personal attribute evaluation dimension of identity: 10% / 22.2% ≈ 45.0%; Adjustment coefficient for the personal attribute evaluation dimension of regional risk: 10% / 16.7% ≈ 59.9%.

[0065] That is to say, the closer the adjustment coefficient is to 1, the closer the importance judgment of the current scoring card model to them is to the risk assessment model, and the smaller the adjustment range required for the scoring card model. Among the above results, the adjustment coefficient for occupation is much greater than 1 (close to 1.5), indicating that in the current scoring card, we underestimated the relative importance of the "occupation" feature. The risk assessment model tells us that from historical data, the actual contribution of this feature in predicting risks is much greater than the weight in the scoring card model.

[0066] Exemplarily, multiply the adjustment coefficient by the corresponding personal attribute score for adjustment to obtain the adjusted personal attribute score, as shown in Table 1:

[0067] In the current embodiment, a first weight is set for the basic risk score, and a second weight is set for the dynamic behavior score. Based on the first weight and the second weight, the basic risk score and the dynamic behavior score are weighted and summed to obtain the risk score of the customer.

[0068] In other embodiments, based on statistics, the risk scores of customers are risk-graded to determine the risk levels of customers. The risk grading steps include: Calculate the current comprehensive risk scores of all customers to form a total sample of scores; Determine the grading thresholds: High-risk threshold: Select the 95th percentile (P95) of the sample. Customers with a score ≥ P95 are defined as high-risk.

[0069] Medium-risk threshold: Select the 75th percentile (P75) of the sample. Customers with a score in the range [P75, P95) are defined as medium-risk.

[0070] Low-risk customers: Customers with a score < P75 are defined as low-risk.

[0071] Advantages: The grading standard changes dynamically according to the overall risk distribution of customers, can automatically adapt to business development, and ensure that risk resources are always focused on the most dangerous customer groups at the top.

[0072] In other embodiments, differentiated risk control is implemented. For low-risk customers, business processes are simplified, and they enjoy fast underwriting and policy maintenance services, with regular spot checks and audits (e.g., once a year). For medium-risk customers, standard due diligence processes are executed, with system alerts for large-amount policy maintenance, policy surrenders, and loan transactions, and frontline staff conduct more inquiries and implement quarterly risk reviews. For high-risk customers, automatic early warning is implemented: the system automatically triggers early warnings and pushes them to the anti-money laundering compliance specialist's workbench; due diligence (EDD) is strengthened: the EDD process is forcibly initiated, requiring customers to provide supplementary materials such as income verification and proof of source of funds; transaction restrictions: the system can automatically restrict them from engaging in high-risk operations such as increasing insurance coverage, large-amount policy loans, and policy surrenders through non-original channels, or force any such operations to undergo a second manual review; and continuous monitoring: they are included in a high-risk customer list and monitored weekly or in real time.

[0073] Example 2 Based on the same concept, referencing Figure 2 This application also proposes a machine learning-based life insurance customer fraud risk assessment device, comprising: The acquisition module is used to acquire the customer's risk information, which includes insurance type information, personal attribute information and insurance change information. Based on the insurance type information, the product risk score is calculated, a pre-trained scoring card model is performed, and personal attribute information is input into the scoring card model. The scoring card model calculates an initial personal attribute risk score based on the personal attribute information based on the preset scoring card weights, and calculates a dynamic behavior score based on the insurance change information. The correction module extracts the model feature weights corresponding to personal attribute information from the pre-trained risk assessment model, uses the ratio of model feature weights to scorecard weights as adjustment coefficients, and adjusts the initial personal attribute risk score using the adjustment coefficients to obtain the corrected personal attribute risk score. In this module, a binary classification architecture is constructed, which is set to predict whether a customer is a money laundering risk user or a non-money laundering risk user based on the customer's risk information. The risk information of multiple historical customers is used as a training dataset, and the binary classification architecture is trained using the training dataset to obtain the trained risk assessment model. The assessment module uses the sum of the product risk score and the modified personal attribute risk score as the base risk score. The base risk score is then weighted and summed with the dynamic behavior score to obtain the customer's risk score.

[0074] Example 3 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, the memory 404 storing a computer program and the processor 402 being configured to run the computer program to perform the steps in any of the above method embodiments.

[0075] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0076] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0077] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0078] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the machine learning-based life insurance customer fraud risk assessment methods in the above embodiments.

[0079] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0080] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0081] Input / output device 408 is used to input or output information. In this embodiment, the input information may be customer risk information, and the output information may be the customer's risk score.

[0082] Optionally, in this embodiment, the processor 402 can be configured to perform the following steps via a computer program: Obtain customer risk information, including insurance type information, personal attribute information, and insurance change information; calculate product risk score based on insurance type information; pre-train a scoring card model; input personal attribute information into the scoring card model; the scoring card model calculates an initial personal attribute risk score based on preset scoring card weights; and calculate a dynamic behavior score based on insurance change information. The model feature weights corresponding to personal attribute information in the pre-trained risk assessment model are extracted. The ratio of the model feature weights to the scorecard weights is used as the adjustment coefficient. The initial personal attribute risk score is adjusted using the adjustment coefficient to obtain the corrected personal attribute risk score. A binary classification architecture is constructed. The binary classification architecture is set to predict whether the insured is a money laundering risk user or a non-money laundering risk user based on the insured's risk information. The risk information of multiple historical insureds is used as the training dataset. The binary classification architecture is trained using the training dataset to obtain the trained risk assessment model. The base risk score is obtained by weighting and summing the product risk score and the modified personal attribute risk score.

[0083] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0084] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0085] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 3 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0086] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for correcting a personal insurance customer fraud risk assessment based on machine learning, characterized by, The method comprises the following steps: obtaining risk information of a customer, the risk information comprising insurance type information, personal attribute information and insurance change information, calculating a product risk score based on the insurance type information, pre-training a scoring card model using risk information of a plurality of historical customers as a training data set, inputting the personal attribute information into the scoring card model, the scoring card model calculating an initial personal attribute risk score based on pre-set scoring card weights, and calculating a dynamic behavior score based on the insurance change information; extracting model feature weights corresponding to the personal attribute information in the pre-trained risk assessment model, using the ratio of the model feature weights to the scoring card weights as an adjustment coefficient, and adjusting the initial personal attribute risk score using the adjustment coefficient to obtain a corrected personal attribute risk score, wherein a binary classification architecture is constructed, the binary classification architecture is configured to predict whether the customer is a money laundering risk user or a non-money laundering risk user based on the risk information of the customer, a plurality of historical customer risk information is used as a training data set, and the binary classification architecture is trained using the training data set to obtain a trained risk assessment model; summing the product risk score and the corrected personal attribute risk score to obtain a basic risk score, and weighting and summing the basic risk score and the dynamic behavior score to obtain a risk score of the customer.

2. The method for assessing the risk of personal insurance customer fraud based on machine learning correction according to claim 1, characterized in that, The insurance type information is basic information of an insurance type that the customer intends to purchase, the insurance type information corresponds to at least one insurance type evaluation dimension, the insurance type information of the customer is obtained in each insurance type evaluation dimension based on a pre-set insurance type scoring rule, and the sum of the insurance type scores of all insurance type evaluation dimensions is taken as the product risk score of the customer.

3. The method of claim 1, wherein the method is characterized by, The personal attribute information corresponds to at least one personal attribute evaluation dimension, the scoring card model maps each personal attribute evaluation dimension to a corresponding personal attribute score based on a pre-set scoring card weight, and the sum of the personal attribute scores of all personal attribute evaluation dimensions is taken as the initial personal attribute risk score of the customer.

4. The method of claim 1, wherein, The insurance change information is dynamic behavior information of the customer in the insurance business, the insurance change information corresponds to at least one insurance change dimension, the insurance change information of the customer is obtained in each insurance change evaluation dimension based on a pre-set insurance change scoring rule, and the sum of the insurance change scores of all insurance change evaluation dimensions is taken as the dynamic behavior of the customer.

5. The method for assessing the risk of personal insurance customer fraud based on machine learning correction according to claim 4, characterized in that, The operation amount is the premium change amount in a set time period, the premium change amount is mapped to an absolute amount score and a relative amount score in the insurance change rule, and the higher value score between the absolute amount score and the relative amount score is taken as the corresponding insurance change score, wherein the absolute amount score is obtained based on the size of the premium change amount, and the relative amount score is obtained based on the ratio of the premium change amount to the total premium.

6. The method of claim 1, wherein, The risk assessment model adopts a LightGBM architecture or an XGBoost architecture, in the training process of the risk assessment model, the money laundering risk users in the training data set are marked as positive samples, and the non-money laundering risk users are marked as negative samples, and the model feature weights of each personal attribute evaluation dimension corresponding to the personal attribute information are extracted after the training is completed.

7. The method of claim 1, wherein the method is based on machine learning correction of personal insurance customer fraud risk assessment. The model feature weight of each personal attribute evaluation dimension corresponding to the personal attribute information is obtained in the pre-trained risk assessment model, the score card weight of each personal attribute evaluation dimension corresponding to the personal attribute information is obtained in the score card model, the ratio of the model feature weight and the score card weight under the corresponding personal attribute evaluation dimension is taken as the adjustment coefficient of the corresponding personal attribute evaluation dimension, the personal attribute score of each personal attribute evaluation dimension is adjusted by the adjustment coefficient to obtain the adjusted personal attribute score, and the sum of the adjusted personal attribute scores of all personal attribute evaluation dimensions is taken as the corrected personal attribute risk score of the customer.

8. A device for correcting a personal insurance customer fraud risk assessment based on machine learning, characterized by, Comprise: The acquisition module is used for acquiring the risk information of the customer, the risk information comprises insurance type information, personal attribute information and insurance change information, a product risk score is calculated based on the insurance type information, a pre-trained score card model is used, the personal attribute information is input into the score card model, the score card model calculates an initial personal attribute risk score based on a preset score card weight, and a dynamic behavior score is calculated based on the insurance change information; The correction module extracts the model feature weight corresponding to the personal attribute information in the pre-trained risk assessment model, takes the ratio of the model feature weight and the score card weight as an adjustment coefficient, adjusts the initial personal attribute risk score by using the adjustment coefficient to obtain a corrected personal attribute risk score, wherein a binary classification architecture is constructed, the binary classification architecture is configured to predict whether the customer is a money laundering risk user or a non-money laundering risk user based on the risk information of the customer, a plurality of historical customer risk information is used as a training data set, the binary classification architecture is trained by using the training data set to obtain a trained risk assessment model; The evaluation module takes the sum of the product risk score and the corrected personal attribute risk score as a basic risk score, and performs weighted summation on the basic risk score and the dynamic behavior score to obtain the risk score of the customer. 9.An electronic device comprising a memory and a processor, the electronic device characterized by, The memory stores a computer program, and the processor is configured to run the computer program to execute the method of claim 1-7.

10. A readable storage medium, characterized by, The readable storage medium stores a computer program, and the computer program is executed by the processor to realize the method of claim 1-7.

Images