Scoring model training method, scoring method and related device

Abstract

The embodiment of the invention discloses a scoring model training method, a scoring method and a related device, and the method comprises the steps: firstly obtaining a training set containing a purchasing power scoring label, and each sample comprises the basic information, asset information, insurance information and insurance dynamic behavior data of a customer; the method comprises the following steps: performing feature engineering processing on a training set sample to generate a plurality of first feature vectors, encoding the first feature vectors into a numerical form to obtain a first feature value, and further constructing a first two-dimensional feature matrix; and processing the matrix through an attention mechanism to obtain a first activation matrix, and training a binary classifier based on the activation matrix and the purchasing power scoring label to finally obtain a scoring model. According to the method, the purchasing power is comprehensively evaluated by fusing the basic information, assets, insurance and dynamic behavior data of the customer, the evaluation accuracy is improved, and the purchasing power prediction is dynamically adjusted according to the real-time demand of the customer and the market change.

Description

Technical Field

[0001] The present application relates to the field of insurance technology, and in particular to a scoring model training method, a scoring method and related devices. Background Art

[0002] In today's insurance industry, as the industry enters a new era of stock game and high-quality transformation, the refined management of customer value has become a core issue in the development of the industry. The traditional "scale effect" is gradually replaced by more personalized and differentiated market strategies, and how to formulate targeted marketing and service strategies through precise customer segmentation has become a general consensus in the industry. In order to achieve this precise customer segmentation, it is particularly important to understand and identify the level of customer purchasing power. The assessment of customer purchasing power level is not only related to the pricing strategy of insurance products, but also directly affects the effective allocation of company resources and the maximization of customer life cycle value.

[0003] However, the existing insurance customers' purchasing power level identification technology mainly relies on manual means and simple static label evaluation methods. Although this method provides preliminary data support for the industry to a certain extent, it has obvious defects, mainly reflected in the low efficiency of data collection and the difficulty in verifying the authenticity of data.

[0004] Therefore, how to improve the accuracy and efficiency of identifying customer purchasing power levels is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] Based on the above problems, the present application provides a scoring model training method, a scoring method and related devices, which can improve the accuracy and efficiency of identifying the level of customer purchasing power. The present application embodiment discloses the following technical solutions:

[0006] A scoring model training method, the method comprising:

[0007] Obtaining a training set; the training set includes a plurality of samples with purchasing power score labels, wherein one of the samples includes a basic information dataset, an asset information dataset, an insurance information dataset, and an insurance dynamic behavior dataset of a sample customer;

[0008] Performing feature engineering processing on each sample in the training set to obtain a plurality of first feature vectors;

[0009] Encoding each of the first eigenvectors into a numerical form to obtain a first eigenvalue, and organizing all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix;

[0010] Performing attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix;

[0011] A binary classifier is trained based on the first activation matrix and the purchasing power rating label to obtain a rating model.

[0012] In a possible implementation, performing attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix includes:

[0013] Performing element-by-element multiplication of the first two-dimensional feature matrix with itself to obtain an attention score matrix, and normalizing the attention score matrix to obtain a normalized matrix;

[0014] The normalized matrix is ​​converted into a probability distribution using a Softmax function to obtain a probability matrix;

[0015] Multiplying the probability matrix and the elements in the first two-dimensional feature matrix correspondingly to obtain a weighted feature matrix;

[0016] The weighted feature matrix is ​​activated by using a nonlinear activation function to obtain the first activation matrix.

[0017] In a possible implementation, feature engineering is performed on the asset information dataset of the sample customers to obtain a plurality of first feature vectors, including:

[0018] For each asset information in the asset information data set, obtain its comprehensive revision rate in the past M years; the comprehensive revision rate is the revision rate after dimensionality reduction through principal component analysis PCA; M is a positive integer;

[0019] Calculate the revised asset price of the asset information based on the original asset price of the asset information and M comprehensive revision rates using an asset revision formula;

[0020] Using the modified asset price as the first feature vector of the asset information;

[0021] The asset correction formula is: P new is the revised asset price; P ori is the original asset price; r n is the comprehensive revision rate for the nth year in the past, where n is less than or equal to M; ∏ represents multiplication.

[0022] In a possible implementation, feature engineering is performed on the insurance dynamic behavior data of sample customers to obtain a plurality of first feature vectors, including:

[0023] According to the behavior time of different dynamic behaviors of customers, the insurance dynamic behavior data is processed into a time series composed of multiple features, wherein the behavior at each time point is described by multiple first feature vectors.

[0024] In a possible implementation, the step of training a binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model includes:

[0025] Inputting the first activation matrix into the binary classifier to obtain an output score;

[0026] Calculating the difference between the output score and the purchasing power score label to obtain a model error;

[0027] Based on the model error, the cross entropy loss function is used to iteratively adjust the parameters of the binary classifier to obtain the scoring model.

[0028] A scoring method, comprising:

[0029] Acquire a scoring data set of target customers; the scoring data set includes a basic information data set, an asset information data set, an insurance information data set, and an insurance dynamic behavior data set of the target customers;

[0030] Performing feature engineering processing on the data in the scoring data set of the target customer to obtain a plurality of second feature vectors;

[0031] Encoding each of the second eigenvectors into a numerical form to obtain a first eigenvalue, and organizing all of the second eigenvalues ​​into a second two-dimensional eigenmatrix;

[0032] Performing attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix;

[0033] The second activation matrix is ​​input into a scoring model to obtain a purchasing power score of the target customer; the scoring model is trained according to the scoring model training method described above.

[0034] A scoring model training device, the method comprising:

[0035] A first acquisition unit is used to acquire a training set; the training set includes a plurality of samples with purchasing power score labels, and one of the samples includes a basic information data set, an asset information data set, an insurance information data set, and an insurance dynamic behavior data set of a sample customer;

[0036] A first feature engineering processing unit, configured to perform feature engineering processing on each sample in the training set to obtain a plurality of first feature vectors;

[0037] A first feature matrix generating unit, used for encoding each of the first feature vectors into a numerical form to obtain a first eigenvalue, and organizing all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix;

[0038] A first attention mechanism processing unit, configured to perform attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix;

[0039] A model training unit is used to perform model training on a binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model.

[0040] A scoring device, the method comprising:

[0041] The second acquisition unit is used to acquire a scoring data set of a target customer; the scoring data set target customer's basic information data set, asset information data set, insurance information data set and insurance dynamic behavior data set;

[0042] A second feature engineering processing unit, configured to perform feature engineering processing on the data in the target customer's scoring data set to obtain a plurality of second feature vectors;

[0043] A second characteristic matrix generating unit, used for encoding each of the second characteristic vectors into a numerical form to obtain a first eigenvalue, and organizing all the second eigenvalues ​​into a second two-dimensional characteristic matrix;

[0044] A second attention mechanism processing unit, configured to perform attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix;

[0045] An input unit is used to input the second activation matrix into a scoring model to obtain a purchasing power score of a target customer; the scoring model is trained according to the scoring model training method described above.

[0046] A scoring model training device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the scoring model training method as described above is implemented.

[0047] A scoring device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the scoring method described above is implemented.

[0048] Compared with the prior art, this application has the following beneficial effects:

[0049] The present application provides a training method, a scoring method and a related device for a scoring model. Specifically, when executing the training of the scoring model provided in the embodiment of the present application, a training set can be first obtained, and the training set includes multiple samples with purchasing power scoring labels, and one sample includes a basic information data set, an asset information data set, an insurance information data set and an insurance dynamic behavior data set of a sample customer. Then, feature engineering processing is performed for each sample in the training set to obtain multiple first eigenvectors. Then each first eigenvector is encoded into a numerical form to obtain a first eigenvalue, and all first eigenvalues ​​of all samples are organized into a first two-dimensional feature matrix. The first two-dimensional feature matrix is ​​processed by an attention mechanism to obtain a first activation matrix. Finally, a binary classifier is trained based on the first activation matrix and the purchasing power scoring label to obtain a scoring model. The present application evaluates purchasing power by fusing multiple data sets (including basic information, asset information, insurance information and insurance dynamic behavior data of the customer). This integration of multidimensional data can not only more comprehensively reflect the actual situation of the customer, but also improve the accuracy and reliability of the evaluation results. At the same time, by utilizing the insurance dynamic behavior data set, the model can capture the changes in the insurance behavior of the customer in different periods. These dynamic data can effectively reflect customers' real-time needs and purchasing intentions, so that the scoring model can not only reflect the customer's current economic situation, but also adjust the forecast of customer purchasing power in a timely manner to adapt to market changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A method flow chart of a scoring model training method provided in an embodiment of the present application;

[0052] Figure 2 A method flow chart of a feature engineering processing method for an asset information data set provided in an embodiment of the present application;

[0053] Figure 3 A method flow chart of an attention mechanism processing method provided in an embodiment of the present application;

[0054] Figure 4 A method flow chart of a scoring method provided in an embodiment of the present application;

[0055] Figure 5 A schematic diagram of the structure of a scoring model training device provided in an embodiment of the present application;

[0056] Figure 6 A schematic diagram of the structure of a scoring device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0058] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.

[0059] In today's insurance industry, as the market enters a new era of stock competition and high-quality transformation, accurate customer value management has become a core issue. The traditional scale effect is gradually replaced by personalized and differentiated market strategies, and customer segmentation and targeted marketing strategies have become an industry consensus. Identifying customer purchasing power levels is the key to achieving accurate segmentation, which not only affects insurance product pricing, but is also directly related to resource allocation and maximizing customer life cycle value. However, current purchasing power evaluation methods mainly rely on manual and static labels, with obvious problems such as inefficient data collection and difficulty in authenticity verification.

[0060] In addition, the current purchasing power evaluation method lacks the ability to capture the changes in assets over a long period of time and the application of dynamic behavior data. The lack of capturing the changes in assets over a long period of time will lead to evaluation distortion. For example, there is a significant difference in purchasing power between 100 yuan ten years ago and 100 yuan now. The lack of application of dynamic behavior data will make it difficult to fully capture the changes in customers' purchasing power levels and the dynamic characteristics of their consumption capacity.

[0061] In order to solve this problem, a training method, a scoring method and a related device for a scoring model are provided in an embodiment of the present application. First, a training set is obtained, in which each sample contains a basic information data set, an asset information data set, an insurance information data set and an insurance dynamic behavior data set of a customer, and is marked with a purchasing power scoring label. Next, feature engineering is performed on each sample to generate multiple first feature vectors, and then these feature vectors are encoded into numerical form and organized into a first two-dimensional feature matrix. Then, the first two-dimensional feature matrix is ​​processed by an attention mechanism to obtain a first activation matrix. Finally, a binary classifier is model trained based on the first activation matrix and the purchasing power scoring label to obtain a scoring model. The present application evaluates purchasing power by fusing multiple data sets (including basic information, asset information, insurance information and insurance dynamic behavior data of the customer), which not only fully reflects the actual situation of the customer and improves the accuracy of the evaluation, but also uses dynamic behavior data to capture the customer's real-time needs and purchasing intentions, and adjusts the forecast in time to adapt to market changes.

[0062] In addition, steps A1-A3 can effectively solve the problem that traditional asset valuation methods cannot capture changes in long-term cycles by using comprehensive correction rates and principal component analysis (PCA) dimensionality reduction. This method not only improves the accuracy of the assessment and reduces human bias, but also enables asset valuation to better adapt to the dynamically changing economic environment, and has high practical value and predictive ability.

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0064] See also Figure 1 , which is a flow chart of a method for training a scoring model provided in an embodiment of the present application, such as Figure 1 As shown, the training of the scoring model may include steps S101-S105:

[0065] S101: Obtain a training set.

[0066] The training set obtained in this step is composed of multiple samples, each of which has a "purchasing power score label", which reflects the purchasing power level of the customer. In this context, the "purchasing power score" is usually evaluated by analyzing various types of customer information to assess their consumption potential and economic ability.

[0067] The data for each sample consists of four main parts:

[0068] Basic information dataset: This part of the data describes the basic situation of the customer, such as the customer's age, gender, family structure, education level, occupation information, etc. This basic information can help evaluate factors such as the customer's socioeconomic background and life stage, which are usually closely related to the customer's purchasing power and insurance needs. For example, older customers may be more concerned about health insurance, while young families may be more concerned about children's education insurance or home property insurance.

[0069] Asset information data set: Asset information includes the customer's car price, house price, regular premium, annuity balance, etc. This data can be used to understand the customer's asset status, as well as their ability to pay and demand for insurance products. For example, customers with higher assets may tend to choose high-end insurance products, while customers with low income or less assets may only choose basic insurance types.

[0070] Insurance information dataset: This part of information covers the customer's insurance-related information, such as the number of policies, insured amount, insurance time, historical compensation, etc. This information helps analyze the customer's past insurance behavior and needs, and then infer their possible insurance needs in the future. For example, if a customer has multiple major disease insurance policies, it may mean that he has a high degree of attention to health insurance, or it may mean that he has a high health risk.

[0071] Insurance dynamic behavior data set: This part of the data contains customer behavior data in different channels, such as interactions through telephone, corporate WeChat, stores, official websites, applications (APP), etc. The behavior data records the details of the customer's interaction time, interaction duration, and interaction insurance types on these channels. For example, if a customer views health insurance-related content on the APP many times, it may indicate that he is more interested in health insurance or intends to purchase health insurance. By analyzing these behavioral data, we can further understand the customer's interests, preferences, and potential needs.

[0072] By collecting multi-dimensional information of customers (including basic information, asset information, insurance information and behavioral data), data support is provided for training models. This data of each sample will be used to evaluate the customer's purchasing power score, thereby helping to build a machine learning model that can predict customer purchasing power. Through the integration and analysis of this information, users can more accurately identify customer needs and provide personalized insurance products or services.

[0073] S102: Perform feature engineering processing on each sample in the training set to obtain multiple first feature vectors.

[0074] Performing feature engineering on each sample in the training set means refining the raw data into a more representative and easier-to-understand feature representation through a series of data transformation and processing steps. In this case, feature engineering may involve processing the customer's basic information, insurance information, asset information, and behavioral data to extract important features or attributes. These processing steps may include missing value processing, feature scaling, feature selection, feature transformation (such as one-hot encoding or standardization), feature synthesis, and other methods to generate multiple first feature vectors that can better describe the customer's characteristics and serve as input for model training to evaluate the customer's purchasing power. Through feature engineering, the model can better understand and utilize the raw data, improving model performance and predictive capabilities.

[0075] In a possible implementation, data preprocessing is required before feature engineering is performed on samples, and then feature engineering is performed on the preprocessed samples. Data preprocessing includes steps such as data cleaning, processing missing values, processing outliers, data conversion, and standardization.

[0076] In one possible implementation, when performing feature engineering on a basic information dataset, it is necessary to extract and transform key information from multiple dimensions to form a more meaningful feature vector. The following is a detailed description of feature engineering for family structure, education level, and occupation information in the basic information dataset:

[0077] Sources of family structure information: Family structure information comes from two aspects: one is the relationship between the insurer and the insured in the policy, and the other is the family information filled in by the customer when taking out insurance.

[0078] How to process family structure information: Based on the customer's marriage and childbearing status, the family structure is classified into multiple subcategories, such as "young single", "young married without children", "young married with children", "middle-aged single", etc. This classification helps to better understand the customer's life cycle stage and possible changes in purchasing power.

[0079] Source of education level information: Education level comes from the education information filled in by the customer when taking out insurance.

[0080] How to process education level information: Categorize education levels into multiple categories, including "high school", "junior college", "undergraduate", "master's degree", "doctoral degree", etc. In this way, you can understand the customer's education level in more detail and use this information to predict their potential purchasing power.

[0081] Sources of occupational information: Occupational information comes from multiple sources, including the size of the employer's company in the group policy (group insurance), the occupational information filled in by the customer when taking out insurance, and the number of times the aviation accident insurance is taken out.

[0082] Processing method of occupational information: Based on the size of the employer's enterprise, generate the characteristics of whether the employee belongs to an enterprise above a certain size; based on the occupational information filled in by the customer and the size of the enterprise, generate the characteristics of whether the employee is the person in charge of the enterprise; combined with the number of aviation accident insurance policies, generate the characteristics of whether the employee is a business traveler.

[0083] Through the above feature engineering processing, the original basic information data set can be converted into a more descriptive and predictive feature vector. These features can not only reflect the customer's family status, education level and occupational characteristics, but also help construct a more accurate model, thereby improving the accuracy of predicting the customer's purchasing power score.

[0084] In one possible implementation, when performing feature engineering on an insurance information dataset, it is necessary to extract and transform key information from multiple dimensions to form a more meaningful feature vector. The following is a detailed description of feature engineering on the insurance type classification and related information in the insurance information dataset:

[0085] (1) Insurance classification:

[0086] Main categories: The insurance types in the insurance information dataset are mainly divided into two categories: life insurance and property insurance.

[0087] Life insurance: further subdivided into financial management type, traditional life insurance, critical illness type, medical type and accident type.

[0088] Property insurance: further subdivided into auto insurance, home insurance and other insurance types.

[0089] (2) Feature generation:

[0090] ① Generation of personal insurance characteristics:

[0091] Financial type: The time difference between the last insurance policy purchase date and the current date: Calculate the time difference between the last insurance policy purchase date and the current date to reflect the customer's recent demand for financial insurance;

[0092] Historical number of financial management insurance policies: Statistics on the number of all financial management insurance policies that customers have purchased in the past, reflecting the overall financial management insurance demand of customers;

[0093] Number of financial management policies purchased in the past year: Statistics on the number of financial management policies purchased by customers in the past year, reflecting the demand trend of financial management insurance in the short term.

[0094] ②Other insurance types (such as traditional life insurance, critical illness insurance, medical insurance and accident insurance):

[0095] Similar features can also be generated, such as "the difference between the last traditional life insurance policy issuance date and the current date", "the number of historical traditional life insurance policies issued", "the number of traditional life insurance policies issued in the past year", etc.

[0096] (3) Generation of property insurance characteristics:

[0097] ① Car insurance:

[0098] The time difference between the last car insurance policy date and the current date: Calculate the time difference between the last car insurance policy date and the current date, reflecting the customer's recent car insurance demand activity.

[0099] Historical auto insurance purchases: Statistics on the number of all auto insurance purchased by the customer in the past, reflecting the customer's overall auto insurance needs;

[0100] Number of auto insurance policies purchased in the past year: Statistics on the number of auto insurance policies purchased by customers in the past year, reflecting the trend of auto insurance demand in the short term.

[0101] ② Home property insurance and other insurance types:

[0102] Similar features can also be generated, such as "the difference between the last home property insurance policy date and the present date", "the number of historical home property insurance policies", "the number of home property insurance policies in the past year", etc.

[0103] Through the above feature engineering processing, the original insurance information data can be converted into feature vectors with more descriptive and predictive capabilities. These features can not only reflect the different insurance needs and insurance behaviors of customers, but also help the model to more accurately predict the purchasing power score of customers. This method helps to better understand the insurance behavior patterns of customers, thereby improving the prediction accuracy of the model.

[0104] When performing feature engineering on the asset information dataset of sample customers, it is necessary to extract and transform key information from multiple dimensions to form a more meaningful feature vector. Figure 2 , Figure 2 A method flow chart of a method for feature engineering processing of an asset information data set provided in an embodiment of the present application, "performing feature engineering processing on the asset information data set of a sample customer to obtain a plurality of first feature vectors" can be specifically implemented by steps A1-A3:

[0105] A1: For each asset information in the asset information data set, obtain its comprehensive revision rate for the past M years.

[0106] When performing feature engineering on the asset information dataset of sample customers, it is necessary to obtain the comprehensive revision rate of each asset information for the past M years. The comprehensive revision rate here is the revision rate after PCA dimensionality reduction, and M is a positive integer.

[0107] Specifically, first, we need to extract the historical revision rate data of each asset from the asset information dataset. This data includes the price changes of the asset in the past M years, where M is a positive integer representing the choice of time span. For example, if M is 5, we need to consider the data of the past 5 years.

[0108] Next, in order to make more effective use of these historical revision rate data, PCA is used to reduce their dimensionality. PCA is a commonly used statistical analysis technique that can transform high-dimensional data into lower-dimensional data while retaining the main features and information. Specifically, after PCA dimensionality reduction, the comprehensive revision rate we obtained not only contains the main trend of the original revision rate, but also reduces the redundancy and noise of the data, making subsequent calculations more accurate and efficient.

[0109] Furthermore, the comprehensive correction rate after dimensionality reduction is applied to the value correction of the asset to obtain the corrected asset price. In this step, an asset correction formula is used to calculate the corrected price of each asset. The formula is usually calculated by combining the original price of the asset and M comprehensive correction rates to reflect the change in the asset value.

[0110] For example, assume that there is the following data of asset A:

[0111] Year (Y1): 2018, original revised rate 2%;

[0112] Year (Y2): 2019, original revised rate 3%;

[0113] Year (Y3): 2020, original revised rate -1%;

[0114] Year (Y4): 2021, original revised rate 4%;

[0115] Year (Y5): 2022, original revised rate 2%.

[0116] Step 1: Get the original correction rate:

[0117] First, extract the raw revision rates for each year from the dataset:

[0118] 2018: 2%; 2019: 3%; 2020: -1%; 2021: 4%; 2022: 2%.

[0119] Step 2: PCA dimensionality reduction:

[0120] Next, we use PCA to reduce the dimensionality of these original correction rates. Assume that after PCA processing, we get the following comprehensive correction rates:

[0121] Comprehensive correction rate 1: 2.5%; Comprehensive correction rate 2: 3.2%; Comprehensive correction rate 3: 1.8%; Comprehensive correction rate 4: 4.0%; Comprehensive correction rate 5: 2.0%.

[0122] It should be noted that asset information usually contains specific data and details describing each asset. These assets can include stocks, bonds, real estate, etc.

[0123] In summary, by obtaining the comprehensive revision rate of each asset information in the asset information data set for the past M years and performing PCA dimensionality reduction processing, a more concise and effective set of feature vectors can be obtained. These feature vectors can better reflect the historical value change trend of the asset and provide more accurate and reliable input for subsequent model predictions. This method not only simplifies the data processing process, but also improves the prediction performance of the model.

[0124] A2: Calculate the revised asset price of the asset information based on the original asset price of the asset information and M comprehensive revision rates using an asset revision formula.

[0125] When performing feature engineering on asset information datasets, it is necessary not only to extract and convert the original data, but also to use these data to calculate new feature values. That is, based on the above original asset price and M comprehensive correction rates, the asset correction formula is used to calculate the corrected asset price.

[0126] The asset correction formula is: P new is the revised asset price; P ori is the original asset price; r n is the comprehensive revision rate for the nth year in the past, where n is less than or equal to M; ∏ represents multiplication.

[0127] Through the asset correction formula, the original asset price can be combined with multiple historical correction rates to reflect the overall value change of the asset over the past M years. This revised asset price not only takes into account the initial value of the asset, but also integrates its price changes in different years.

[0128] A3: Use the revised asset price as the first feature vector of the asset information.

[0129] Finally, the calculated corrected asset price is used as the first eigenvector of the asset. This eigenvector can more accurately describe the historical value changes and corrections of the asset, which is helpful for subsequent model training and prediction.

[0130] By using the asset correction formula to calculate the revised asset price based on the original asset price and M comprehensive correction rates of asset information, more accurate and meaningful feature vectors can be generated. These feature vectors can better reflect the trend of asset value changes, provide richer input information for the model, and thus improve the prediction accuracy of the model.

[0131] In one possible implementation, when performing feature engineering on the insurance dynamic behavior data of sample customers, we first need to organize the different behaviors of customers into an orderly behavior sequence in chronological order. Specifically, various historical records of customers (such as browsing records, consultation records, and insurance records) can be sorted in time to form a coherent behavior sequence. Each behavior record contains multiple features, such as behavior content, behavior time, length of stay, and related insurance types.

[0132] In this way, the customer's behavior data can be processed into a time series consisting of multiple features. In this time series, the behavior at each time point is described by multiple first feature vectors. These feature vectors contain detailed information about the behavior, making it possible to more fully understand the customer's behavior patterns and preferences.

[0133] For example, suppose a customer has the following behaviors in the past period of time:

[0134] (1) Browsing history:

[0135] Behavior content: browsing the car insurance page;

[0136] Action time: 2023-01-0110:00;

[0137] Duration: 5 minutes;

[0138] Related insurance types: auto insurance.

[0139] (2) Consultation records:

[0140] Behavior: Calling customer service to inquire about auto insurance;

[0141] Behavior time: 2023-01-02 14:00;

[0142] Duration: 10 minutes;

[0143] Related insurance types: auto insurance.

[0144] (3) Insurance record:

[0145] Behavior content: Submitting a car insurance application;

[0146] Action time: 2023-01-0316:00;

[0147] Duration: 20 minutes;

[0148] Related insurance types: auto insurance.

[0149] By sorting these behaviors by time, an ordered behavior sequence is formed, and each behavior record is converted into a first feature vector containing multiple features. For example, it can be expressed as:

[0150] Time point 1: [browsing car insurance page, 2023-01-01 10:00, 5 minutes, car insurance];

[0151] Time point 2: [Call customer service to inquire about auto insurance, 2023-01-02 14:00, 10 minutes, auto insurance];

[0152] Time point 3: [Submit car insurance application, 2023-01-0316:00, 20 minutes, car insurance].

[0153] In this way, we can clearly see the customer's behavior trajectory and behavior pattern, so as to better analyze the customer's preferences and needs. This time series-based feature engineering method can not only reflect the customer's dynamic behavior pattern, but also help to more accurately predict future customer behavior and provide customers with more personalized and precise services.

[0154] In summary, through feature engineering of customer insurance dynamic behavior data, not only can the complex behavior data be converted into an ordered behavior sequence and further refined into a first feature vector composed of multiple features, but it can also provide strong support for subsequent behavior analysis and prediction.

[0155] S103: Encode each of the first eigenvectors into a numerical form to obtain a first eigenvalue, and organize all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix.

[0156] In order to further process these feature vectors, each first feature vector needs to be encoded into a numerical form to obtain the first feature value. This process involves converting non-numerical features (such as text, categories, etc.) into numerical representations. For example, one-hot encoding can be used to convert categorical variables into binary vectors, or numerical methods can be used to convert timestamps into specific numerical values.

[0157] The specific steps are as follows:

[0158] Feature encoding: Encode the non-numeric features in each first feature vector into numeric form. For example, the gender ("male" or "female") in the basic information of the customer can be encoded as 0 and 1; the timestamp can be converted into a Unix timestamp. In this way, each first feature vector can be converted into a vector composed of numeric values.

[0159] Exemplarily, it is assumed that a first eigenvector contains the following information:

[0160] Customer name: Zhang San; Age: 35; Gender: Male; Total assets: 500,000; Transaction time: 2023-01-0110:00.

[0161] After encoding, it can be converted into: Name: [1,0,0] (assuming Zhang San is encoded as [1,0,0]); Age: 35; Gender: 0; Total assets: 500000; Timestamp: 1672569600.

[0162] Therefore, the first eigenvector can be expressed as: [1, 0, 0, 35, 0, 500000, 1672569600].

[0163] Organize feature matrix: Organize all first eigenvalues ​​of all samples into a first two-dimensional feature matrix. The first eigenvector of each sample constitutes a row of the matrix, and the eigenvectors of all samples together constitute all rows of the matrix.

[0164] For example, if there are three samples, their first eigenvectors are:

[0165] Sample 1: [1,0,0,35,0,500000,1672569600];

[0166] Sample 2: [0,1,0,40,1,600000,1672656000];

[0167] Sample 3: [0,0,1,45,1,700000,1672742400].

[0168] Then the first two-dimensional feature matrix is ​​as follows:

[0169]

[0170] In this way, the multi-dimensional complex feature data is converted into numerical form and organized into a structured two-dimensional matrix, which is convenient for subsequent data processing and model training. This process can effectively extract and represent important features in the data, laying the foundation for further attention mechanism processing and model training.

[0171] S104: Performing attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix.

[0172] Applying the attention mechanism to the first two-dimensional feature matrix, the first activation matrix is ​​finally obtained, which means that the attention mechanism is introduced in the process of building the scoring model in order to pay more attention to and highlight the key features that affect purchasing power. The attention mechanism can dynamically learn the importance of each feature and adjust the weight of the feature according to its contribution to the purchasing power score, so that the model can focus more on the most decisive features. In this way, the model can more effectively capture relevant patterns and information in customer data and improve the model's ability to accurately predict purchasing power. Therefore, by introducing the attention mechanism to process the feature matrix, the scoring model can be made more intelligent, precise and targeted, thereby improving the performance and effect of the overall model.

[0173] See also Figure 3 , Figure 3 A method flow chart of an attention mechanism processing method provided in an embodiment of the present application. Accordingly, the first two-dimensional feature matrix is ​​processed by the attention mechanism to obtain a first activation matrix, which can be specifically implemented through steps B1-B4:

[0174] B1: Perform element-by-element multiplication on the first two-dimensional feature matrix to obtain an attention score matrix, and normalize the attention score matrix to obtain a normalized matrix.

[0175] The first two-dimensional feature matrix is ​​multiplied element by element by itself to obtain the attention score matrix, and then the obtained attention score matrix is ​​normalized to obtain the normalized matrix. This process is actually to determine the importance and influence of each feature in the model. The self-element-by-element multiplication calculates the attention score by measuring the correlation and association between different features, thereby revealing the interaction between each feature. Further normalization of the attention score matrix can ensure that the weight of each feature is within a reasonable range, making these weights easier to interpret and compare. This processing method helps to accurately locate key features and provides a basis for subsequent steps, so that the model can more accurately capture the important features of the data, providing important support for the establishment of the final scoring model. The whole process aims to optimize the learning and decision-making process of the model so as to better understand and utilize the information in the data, thereby improving the performance and predictive ability of the model.

[0176] For example, for a vector A=[2,3], element-wise multiplication with itself can be expressed as:

[0177] A⊙A=[2,3]⊙[2,3]A⊙A=[2,3]⊙[2,3].

[0178] The calculation process is as follows:

[0179] A⊙A=[2×2,3×3]=[4,9]A⊙A=[2×2,3×3]=[4,9].

[0180] Therefore, the result of the element-wise multiplication is:

[0181] A⊙A=[4,9]A⊙A=[4,9].

[0182] B2: Use the Softmax function to convert the normalized matrix into a probability distribution to obtain a probability matrix.

[0183] The purpose of using the Softmax function to convert the normalized matrix into a probability distribution is to map the original data to a range between 0 and 1, and the sum of all elements is 1, representing the relative importance of different features or items in the overall probability distribution. The Softmax function performs an exponential operation on the input vector, and then divides each element by the sum of the exponentials of all elements to obtain the normalized probability value.

[0184] This conversion can help the model better understand and utilize the weights of the attention scores, so that important features can be highlighted and receive appropriate attention. Through the Softmax function, we distribute attention to different features so that the model can learn and utilize these feature information more effectively, further improving the accuracy and effect of the model. Therefore, converting the normalized matrix into a probability distribution to obtain a probability matrix provides the model with an interpretable and comparable way to measure the importance and contribution of each feature, thereby promoting the model's in-depth understanding and optimization of the data.

[0185] B3: Multiply the probability matrix and the elements in the first two-dimensional feature matrix correspondingly to obtain a weighted feature matrix.

[0186] Multiplying the probability matrix and the elements in the first two-dimensional feature matrix correspondingly to obtain the weighted feature matrix is ​​actually weighting different features and adjusting their influence in the model according to their overall importance. This step is to highlight those features that are considered to have a greater contribution to the final prediction results, so that the model can focus more on learning and utilizing these key information.

[0187] Exemplarily, assume there is a probability matrix P and a first two-dimensional feature matrix F. The dimensions of these two matrices are the same, both 2×3.

[0188] Probability Matrix The first two-dimensional characteristic matrix

[0189] Step 1: Element-wise multiplication

[0190] Multiply the corresponding elements in the probability matrix P and the feature matrix F to generate a new matrix W.

[0191] For P 11 and F 11 :W 11 =P 11 ×F 11 =0.1×5=0.5;

[0192] For P 12 and F 12 :W 12 =P 12 ×F 12 =0.2×10=2.0;

[0193] For P 13 and F 13 :W 13 =P 13 ×F 13 =0.7×15=10.5;

[0194] For P 21 and F 21 :W 21 =P 21 ×F 21 =0.4×20=8.0;

[0195] For P 22 and F 22 :W 22 =P 22 ×F 22 =0.3×25=7.5;

[0196] For P 23 and F 23 :W 23 =P 23 ×F 23 =0.3×30=9.0.

[0197] Through the above calculation, the weighted feature matrix W is obtained:

[0198]

[0199] In this way, a weighted feature matrix W is obtained, in which each element is the product of the corresponding position element of the probability matrix P and the feature matrix F. This weighted feature matrix reflects the weight of each eigenvalue after probability adjustment, which can be used in subsequent machine learning model training or data analysis tasks.

[0200] By multiplying the probability matrix by the original feature matrix element by element, each eigenvalue can be weighted according to its corresponding probability, so that the model focuses more on features with higher weights. This weighted processing helps to improve the model's sensitivity and expressiveness to the data, thereby improving the model's generalization ability and prediction accuracy. Therefore, the process of generating a weighted feature matrix provides the model with a more targeted and optimized feature representation, allowing the model to more effectively capture key information and patterns in the data, thereby improving the performance and effect of the overall model.

[0201] B4: Activate the weighted feature matrix using a nonlinear activation function to obtain the first activation matrix.

[0202] By using a nonlinear activation function to activate the weighted feature matrix, the first activation matrix is ​​obtained. This step is to introduce nonlinear factors and enhance the expressiveness of the model so that the model can better learn complex data patterns and relationships. Nonlinear activation functions play a vital role in neural networks. They can introduce nonlinear transformations, thereby enabling neural networks to fit various nonlinear and complex functions.

[0203] By applying a nonlinear activation function, mapping the results of the weighted feature matrix to another space, more feature information and patterns may be stimulated. This process helps to extract and highlight deeper feature representations in the data, enhance the nonlinear modeling capabilities of the model, and better capture the complex associations and patterns in the data. The final first activation matrix will contain richer and more abstract feature representations, providing a more comprehensive and effective input for the model, enabling the model to make more accurate and reliable decisions in prediction tasks. Therefore, applying a nonlinear activation function to activate the weighted feature matrix plays a key role in optimizing the performance of the model and promoting its learning process, which helps to improve the generalization ability and prediction accuracy of the model.

[0204] S105: Performing model training on a binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model.

[0205] Model training of a binary classifier based on the first activation matrix and the purchasing power score label is a key process, aiming to build a scoring model that can accurately predict the purchasing power score. This process includes several key steps, among which the first activation matrix, as an important feature representation after feature extraction, processing and nonlinear transformation, will be input into the binary classifier to obtain the output score.

[0206] Specifically, first, the first activation matrix is ​​input into the binary classifier, and the output score of the model is calculated through the network's forward propagation. This output score is compared with the purchasing power score label to calculate the model error. This error represents the deviation and accuracy of the model in predicting the purchasing power score. We hope to continuously reduce this error through the training process so that the model's prediction results are closer to the true label value.

[0207] Based on the calculated model error, the cross entropy loss function can be used to quantify the model's prediction error, and then the parameters of the binary classifier can be adjusted and optimized to minimize the loss function, thereby obtaining a more accurate scoring model. Through the back propagation algorithm, the model will fine-tune the parameters according to the error signal, so that the model's prediction results on the training data gradually approach the true label, improving the model's generalization ability and prediction accuracy.

[0208] This training process is an iterative optimization process that continuously adjusts the model parameters to minimize the loss function to improve the performance and accuracy of the model. The final scoring model will be able to better predict the purchasing power score and provide reliable data support for corporate decision-making, helping to optimize market strategies, improve user experience and promote business growth. This process based on feature extraction, model training and parameter adjustment is a key step in building an efficient scoring model. It can effectively deal with classification and prediction problems in the real world and provide a solid foundation for data-driven decision-making and business development.

[0209] Based on the contents of S101-S105, it can be known that a training set containing multiple samples with purchasing power score labels is first obtained, and each sample includes the customer's basic information, asset information, insurance information and dynamic behavior data. Then, feature engineering is performed on each sample to obtain multiple first eigenvectors, and they are encoded into first eigenvalues ​​in numerical form and organized into a first two-dimensional feature matrix. Then the first two-dimensional feature matrix is ​​processed by the attention mechanism to obtain the first activation matrix. Finally, the first activation matrix and the purchasing power score label are used to train the binary classifier model to obtain a scoring model. This application evaluates purchasing power by integrating multiple data sets (including the customer's basic information, asset information, insurance information and insurance dynamic behavior data). This integration of multidimensional data can not only more comprehensively reflect the customer's actual situation, but also improve the accuracy and reliability of the evaluation results. By using the insurance dynamic behavior data set, the model can capture the customer's insurance behavior changes in different periods. These dynamic data can effectively reflect the customer's real-time needs and purchase intentions, so that the scoring model can not only reflect the customer's current economic situation, but also adjust the prediction of the customer's purchasing power in a timely manner to adapt to market changes.

[0210] Based on the above-mentioned embodiment of the training method of the scoring model, the present application embodiment also provides a scoring method. Figure 4 , Figure 4 A method flow chart of a scoring method provided in an embodiment of the present application. Figure 4 As shown, the method includes S401-S405:

[0211] S401: Obtain a rating dataset of target customers.

[0212] In order to evaluate the purchasing power of target customers, we first need to obtain the target customer scoring dataset, which includes the target customer's basic information dataset, asset information dataset, insurance information dataset, and insurance dynamic behavior dataset. These datasets will provide key information and features for the subsequent prediction process.

[0213] S402: Perform feature engineering processing on the data in the target customer's rating data set to obtain a plurality of second feature vectors.

[0214] The data in the target customer rating data set is processed by feature engineering, and multiple second feature vectors are obtained through technical means such as data cleaning, feature selection, and feature conversion. This step aims to prepare data and extract features that are meaningful for purchasing power score prediction, laying the foundation for subsequent model training.

[0215] It should be noted that the specific processing process of step S402 is similar to the specific processing process of step S102, and will not be repeated here.

[0216] S403: Encode each of the second eigenvectors into a numerical form to obtain a first eigenvalue, and organize all of the second eigenvalues ​​into a second two-dimensional eigenmatrix.

[0217] Each second eigenvector is encoded into a numerical form to obtain the first eigenvalue, and all second eigenvalues ​​are organized into a second two-dimensional feature matrix. This matrix will contain the encoded and integrated feature information, providing a complete feature representation for the model input.

[0218] It should be noted that the specific processing process of step S403 is similar to the specific processing process of step S103, and will not be repeated here.

[0219] S404: Performing attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix.

[0220] The second two-dimensional feature matrix is ​​processed by the attention mechanism, and the second activation matrix is ​​obtained by emphasizing the important features that the model pays attention to. This attention mechanism processing helps to strengthen the model's attention to features and weight adjustment, and improve the model's learning efficiency and generalization ability.

[0221] It should be noted that the specific processing process of step S404 is similar to the specific processing process of step S104, and will not be repeated here.

[0222] S405: Input the second activation matrix into a scoring model to obtain a purchasing power score of the target customer.

[0223] Finally, the second activation matrix is ​​input into the scoring model, and the target customer's purchasing power score is obtained according to the trained scoring model. This scoring model is obtained according to the aforementioned scoring model training method. In the prediction process, it can effectively infer the customer's purchasing power status and provide an important reference for business decision-making and marketing.

[0224] Based on the contents of S401-S405, we can see that through multi-dimensional data fusion, advanced feature engineering, attention mechanism and training-based scoring model, we can not only accurately evaluate the purchasing power of customers, but also have strong flexibility and adaptability, and can be continuously optimized.

[0225] See also Figure 5 , Figure 5 A schematic diagram of the structure of a scoring model training device provided in an embodiment of the present application. Figure 5 As shown, the training device of the scoring model includes:

[0226] The first acquisition unit 501 is used to acquire a training set; the training set includes a plurality of samples with purchasing power score labels, and one of the samples includes a basic information dataset, an asset information dataset, an insurance information dataset, and an insurance dynamic behavior dataset of a sample customer;

[0227] A first feature engineering processing unit 502, configured to perform feature engineering processing on each sample in the training set to obtain a plurality of first feature vectors;

[0228] A first feature matrix generating unit 503 is used to encode each of the first feature vectors into a numerical form to obtain a first eigenvalue, and organize all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix;

[0229] A first attention mechanism processing unit 504 is used to perform attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix;

[0230] The model training unit 505 is used to perform model training on the binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model.

[0231] In a possible implementation, the first attention mechanism processing unit 504 specifically includes:

[0232] an element-by-element multiplication calculation unit, configured to perform element-by-element multiplication calculation on the first two-dimensional feature matrix to obtain an attention score matrix, and to normalize the attention score matrix to obtain a normalized matrix;

[0233] A probability matrix generating unit, used for converting the normalized matrix into a probability distribution using a Softmax function to obtain a probability matrix;

[0234] A weighted feature matrix generating unit, used for correspondingly multiplying the probability matrix and the elements in the first two-dimensional feature matrix to obtain a weighted feature matrix;

[0235] An activation matrix generating unit is used to perform an activation operation on the weighted feature matrix using a nonlinear activation function to obtain the first activation matrix.

[0236] In a possible implementation, the first feature engineering processing unit 502 specifically includes:

[0237] A comprehensive correction rate acquisition unit is used to acquire the comprehensive correction rate of each asset information in the asset information data set for the past M years; the comprehensive correction rate is the correction rate after dimensionality reduction through principal component analysis PCA; M is a positive integer;

[0238] A revised asset price calculation unit, used to calculate the revised asset price of the asset information based on the original asset price of the asset information and M comprehensive revision rates using an asset revision formula;

[0239] A setting unit, configured to use the revised asset price as a first feature vector of the asset information;

[0240] The asset correction formula is: P new is the revised asset price; P ori is the original asset price; r n is the comprehensive revision rate for the nth year in the past, where n is less than or equal to M; ∏ represents multiplication.

[0241] In a possible implementation, the first feature engineering processing unit 502 further includes:

[0242] The time series generating unit is used to process the insurance dynamic behavior data into a time series composed of multiple features according to the behavior time of different dynamic behaviors of customers, wherein the behavior at each time point is described by multiple first feature vectors.

[0243] In a possible implementation, the model training unit 505 specifically includes:

[0244] an output score obtaining unit, configured to input the first activation matrix into the binary classifier to obtain an output score;

[0245] An error unit calculation unit, used to calculate the difference between the output score and the purchasing power score label to obtain a model error;

[0246] A parameter adjustment unit is used to iteratively adjust the parameters of the binary classifier based on the model error using a cross entropy loss function to obtain the scoring model.

[0247] See also Figure 6 , Figure 6 This is a schematic diagram of the structure of a scoring device provided in an embodiment of the present application. Figure 6 As shown, the scoring device comprises:

[0248] The second acquisition unit 601 is used to acquire a scoring data set of a target customer; the scoring data set target customer's basic information data set, asset information data set, insurance information data set and insurance dynamic behavior data set;

[0249] A second feature engineering processing unit 602 is used to perform feature engineering processing on the data in the target customer's scoring data set to obtain a plurality of second feature vectors;

[0250] A second characteristic matrix generating unit 603 is used to encode each of the second characteristic vectors into a numerical form to obtain a first eigenvalue, and organize all the second eigenvalues ​​into a second two-dimensional characteristic matrix;

[0251] A second attention mechanism processing unit 604 is used to perform attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix;

[0252] The input unit 605 is used to input the second activation matrix into a scoring model to obtain a purchasing power score of the target customer; the scoring model is trained according to the scoring model training method as described above.

[0253] In addition, an embodiment of the present application also provides a scoring model training device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the scoring model training method described above is implemented.

[0254] In addition, an embodiment of the present application further provides a scoring device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the scoring method described above is implemented.

[0255] The embodiment of the present application provides a training method, a scoring method and related devices for a scoring model. The construction process of the material aging damage grading model includes: obtaining a training set containing a purchasing power scoring label, wherein each sample contains the customer's basic information, asset information, insurance information and insurance dynamic behavior data, and performing feature engineering processing to obtain a first eigenvector. Subsequently, each vector is encoded into a numerical form to obtain a first eigenvalue, and the first eigenvalues ​​of all samples are integrated to form a two-dimensional feature matrix, which is then processed using an attention mechanism to form a first activation matrix. Finally, model training is performed based on the first activation matrix and the purchasing power scoring label to establish a scoring model. The present application comprehensively evaluates purchasing power and improves evaluation accuracy by integrating basic customer information, assets, insurance and dynamic behavior data, and dynamically adjusts purchasing power forecasts based on real-time customer needs and market changes.

[0256] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. Some or all of them can be selected according to actual needs to achieve the purpose of the embodiment scheme. Ordinary technicians in this field can understand and implement it without paying creative work.

[0257] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A scoring model training method, characterized in that: The method comprises: Obtaining a training set; the training set includes a plurality of samples with purchasing power score labels, wherein one of the samples includes a basic information dataset, an asset information dataset, an insurance information dataset, and an insurance dynamic behavior dataset of a sample customer; Performing feature engineering processing on each sample in the training set to obtain a plurality of first feature vectors; Encoding each of the first eigenvectors into a numerical form to obtain a first eigenvalue, and organizing all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix; Performing attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix; A binary classifier is trained based on the first activation matrix and the purchasing power rating label to obtain a rating model.

2. The method according to claim 1, characterized in that The performing attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix includes: Performing element-by-element multiplication of the first two-dimensional feature matrix with itself to obtain an attention score matrix, and normalizing the attention score matrix to obtain a normalized matrix; The normalized matrix is ​​converted into a probability distribution using a Softmax function to obtain a probability matrix; Multiplying the probability matrix and the elements in the first two-dimensional feature matrix correspondingly to obtain a weighted feature matrix; The weighted feature matrix is ​​activated by using a nonlinear activation function to obtain the first activation matrix.

3. The method according to claim 1, characterized in that The asset information dataset of the sample customers is processed by feature engineering to obtain multiple first feature vectors, including: For each asset information in the asset information data set, obtain its comprehensive revision rate in the past M years; the comprehensive revision rate is the revision rate after dimensionality reduction through principal component analysis PCA; M is a positive integer; Calculate the revised asset price of the asset information based on the original asset price of the asset information and M comprehensive revision rates using an asset revision formula; Using the modified asset price as the first feature vector of the asset information; The asset correction formula is: P new is the revised asset price; P ori is the original asset price; r n is the comprehensive revision rate for the nth year in the past, where n is less than or equal to M; ∏ represents multiplication.

4. The method according to claim 1, characterized in that The insurance dynamic behavior data of sample customers is processed by feature engineering to obtain multiple first feature vectors, including: According to the behavior time of different dynamic behaviors of customers, the insurance dynamic behavior data is processed into a time series composed of multiple features, wherein the behavior at each time point is described by multiple first feature vectors.

5. The method according to claim 1, characterized in that The step of training a binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model includes: Inputting the first activation matrix into the binary classifier to obtain an output score; Calculating the difference between the output score and the purchasing power score label to obtain a model error; Based on the model error, the cross entropy loss function is used to iteratively adjust the parameters of the binary classifier to obtain the scoring model.

6. A scoring method, characterized in that: The method comprises: Acquire a scoring data set of target customers; the scoring data set includes a basic information data set, an asset information data set, an insurance information data set, and an insurance dynamic behavior data set of the target customers; Performing feature engineering processing on the data in the scoring data set of the target customer to obtain a plurality of second feature vectors; Encoding each of the second eigenvectors into a numerical form to obtain a first eigenvalue, and organizing all of the second eigenvalues ​​into a second two-dimensional eigenmatrix; Performing attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix; The second activation matrix is ​​input into a scoring model to obtain a purchasing power score of the target customer; the scoring model is trained according to the scoring model training method according to any one of claims 1 to 5.

7. A scoring model training device, characterized in that: The method comprises: A first acquisition unit is used to acquire a training set; the training set includes a plurality of samples with purchasing power score labels, and one of the samples includes a basic information data set, an asset information data set, an insurance information data set, and an insurance dynamic behavior data set of a sample customer; A first feature engineering processing unit, configured to perform feature engineering processing on each sample in the training set to obtain a plurality of first feature vectors; A first feature matrix generating unit, used for encoding each of the first feature vectors into a numerical form to obtain a first eigenvalue, and organizing all the first eigenvalues ​​of all samples into a first two-dimensional feature matrix; A first attention mechanism processing unit, configured to perform attention mechanism processing on the first two-dimensional feature matrix to obtain a first activation matrix; A model training unit is used to perform model training on a binary classifier based on the first activation matrix and the purchasing power rating label to obtain a rating model.

8. A scoring device, characterized in that: The method comprises: The second acquisition unit is used to acquire a scoring data set of a target customer; the scoring data set includes a basic information data set, an asset information data set, an insurance information data set, and an insurance dynamic behavior data set of the target customer; A second feature engineering processing unit, configured to perform feature engineering processing on the data in the target customer's scoring data set to obtain a plurality of second feature vectors; A second characteristic matrix generating unit, used for encoding each of the second characteristic vectors into a numerical form to obtain a first eigenvalue, and organizing all the second eigenvalues ​​into a second two-dimensional characteristic matrix; A second attention mechanism processing unit, configured to perform attention mechanism processing on the second two-dimensional feature matrix to obtain a second activation matrix; An input unit is used to input the second activation matrix into a scoring model to obtain a purchasing power score of a target customer; the scoring model is trained according to the scoring model training method according to any one of claims 1 to 5.

9. A scoring model training device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the training method of the scoring model as described in any one of claims 1 to 5 is implemented.

10. A scoring device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the scoring method according to claim 6 is implemented.

Images