Regression model-based target customer transaction probability prediction method and equipment

Abstract

The invention discloses a target customer transaction probability prediction method and device based on a regression model, and the method comprises the steps: S1, carrying out the dirty data elimination and unstructured data conversion of the original data of a customer of a life insurance company, and obtaining the standardized data; s2, dividing the standardized data into a low-net-value customer data source and a high-net-value customer data source based on the customer income level for subsequent prediction; s3, multiple machine learning models are selected as candidate models, a dual-drive strategy is adopted to screen customer transaction key features, and an initial prediction model is constructed; s4, training the initial prediction model, and determining an optimal parameter through regularization processing and cross validation to obtain an optimized regression prediction model; and S5, performing grouping verification and actual scene verification on the optimized regression prediction model, determining an optimal deal probability threshold, and establishing a long-acting iteration mechanism continuous optimization model at the same time.

Description

Technical Field

[0001] This application relates to the technical field, and in particular to a method and device for predicting the probability of a transaction with a target customer based on a regression model. Background Technology

[0002] In the insurance industry, the existing customer base for life insurance is large, but the number of agents is decreasing year by year, making it difficult to conduct comprehensive visits and business operations. Furthermore, due to the departure of some agents and the reassignment of their "orphans," the newly assigned agents are unfamiliar with the customers, making customer development difficult.

[0003] Life insurance companies' various business activities revolve around customer visits, but they have never clearly defined which customers to prioritize. Over time, partners' willingness to make such visits has gradually decreased. There is an imbalance between reach and conversion: a large number of customers are pushed to them each month, but business partners lack sufficient understanding of these customers, resulting in a mismatch between reach and conversion rates (especially for low-premium customers). There is also an imbalance in attention: low-net-worth customers account for over 95%, but the visit coverage rate is only 11.1%, far lower than the 46.5% for high-net-worth customers, indicating extremely low development. Furthermore, there are limitations in manual analysis: there are more than 20 customer transaction-related characteristics (such as star rating, age, and payment history), making it difficult for manual cross-validation to identify key characteristics and find target customer groups with high conversion rates.

[0004] Therefore, there is an urgent need for a method to predict the probability of closing a deal with a target customer based on a regression model, so as to help sales personnel identify high-quality customers with a high conversion rate. Summary of the Invention

[0005] This application provides a method and device for predicting the probability of a transaction with a target customer based on a regression model, in order to solve the above-mentioned problems.

[0006] On the one hand, this application provides a method for predicting the probability of a transaction with a target customer based on a regression model, the method comprising the following steps: Step S1: Remove dirty data and transform unstructured data from the raw customer data of the life insurance company to obtain standardized data; Step S2: Based on customer income levels, the standardized data is divided into low-net-worth customer data sources and high-net-worth customer data sources, which are used for subsequent predictions respectively; Step S3: Select multiple machine learning models as candidate models, use a dual-drive strategy to screen key features of customer transactions, and build an initial prediction model; Step S4: Train the initial prediction model, determine the optimal parameters through regularization and cross-validation, and obtain the optimized regression prediction model; Step S5: Perform group validation and real-world scenario validation on the optimized regression prediction model to determine the optimal transaction probability threshold, and establish a long-term iterative mechanism to continuously optimize the model.

[0007] In one implementation of this application, step S1, dirty data removal specifically involves: unifying the time format to a preset standard format and removing time data that is not in the standard format; and removing ID number data that is in a preset hash value format.

[0008] In one implementation of this application, in step S1, the unstructured data conversion adopts a tag encoding method to convert Chinese character data into a computer-recognizable digital format, wherein a preset affirmative expression is replaced with a first number, and a preset negative expression is replaced with a second number.

[0009] In one implementation of this application, in step S2, the criteria for dividing customer data sources are as follows: based on the customer star rating system, customers in the low star rating range are defined as low-net-worth customers, and customers in the high star rating range are defined as high-net-worth customers.

[0010] In one implementation of this application, in step S3, the candidate models include a random forest model, a gradient boosting decision tree model, and a logistic regression model; the key features for selecting customer transactions are: if the conversion rate increases when the number of targeted customers decreases, the model is deemed effective, and finally, the model with the best conversion rate under the same number of targeted customers is selected as the initial prediction model.

[0011] In one implementation of this application, in step S3, the dual-drive strategy specifically involves: on the one hand, identifying the implicit characteristics of high-value customers through feedback from sales business partners; on the other hand, conducting data-driven feature importance assessment based on feature importance assessment methods. The feature importance assessment methods include SHAP value analysis or regularized feature selection. The key features selected include near-term customers, customer age, customer gender, expired customers, claim customers, renewal customers, customer annual income, universal account balance, cumulative annual payment for the year, and recent purchase of wealth management products.

[0012] In one implementation of this application, in step S4, the regularization process uses L1 regularization, and the cross-validation is multi-fold cross-validation; the optimal parameters of the model are determined through the regularization process and cross-validation to improve the model accuracy and avoid overfitting; the parameters include: regularization coefficient, number of iterations, learning rate, and classification threshold.

[0013] In one implementation of this application, in step S5, the long-term iterative mechanism specifically involves: tracking customer development process data in real time to build a dynamic evaluation system for model performance, regularly collecting feedback information from the sales team, and continuously optimizing the feature weight matrix and decision threshold through incremental learning and model retraining to alleviate the model drift problem.

[0014] In one implementation of this application, in step S5, the group verification includes verification of past data and determination of the optimal threshold, wherein the optimal threshold is determined by gradient testing; the actual scenario verification is continuous tracking and observation for a preset period, and high-potential customers are selected based on the model and their conversion rate is statistically analyzed.

[0015] On the other hand, this application also provides a device for predicting the probability of a transaction with a target customer based on a regression model. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the aforementioned method for predicting the probability of a transaction with a target customer based on a regression model.

[0016] This application provides a method and device for predicting the probability of a transaction with a target customer based on a regression model, which has the following beneficial effects: 1. Improve data quality and predictive relevance to lay the foundation for accurate forecasting. This method cleans up dirty data by standardizing time formats and removing hash-formatted ID numbers, converts unstructured Chinese character data into numerical format using tag encoding, and categorizes customers into low-net-worth and high-net-worth groups based on a star rating system for separate forecasting. This avoids the singular interference of customer purchasing power in the forecast, ensuring that the data input to the model is standardized and of high quality, making subsequent forecasts more closely aligned with the characteristics of different customer groups.

[0017] 2. Optimize model performance to overcome the limitations of manual analysis and accurately identify high-conversion-rate customers. The method employs a dual-drive strategy to select 10 key features, avoiding the difficulty of cross-validating more than 20 features manually. Optimal parameters (e.g., accuracy up to 91%) are determined through L1 regularization and multi-fold cross-validation. The model with the highest conversion rate among the same number of selected customers is then selected, effectively improving prediction accuracy, avoiding overfitting, and efficiently identifying high-conversion-rate, high-quality customers.

[0018] 3. The model is thoroughly validated and adapted to business dynamics, significantly improving the efficiency of life insurance sales. Group validation (including validation using past data and gradient testing to determine the optimal threshold) and continuous, pre-set periodic validation in real-world scenarios (e.g., three months) ensure the model's effectiveness in real-world business (a 30% conversion rate for labeled customers, 17 percentage points higher than unlabeled customers, resulting in a 131% increase in conversion rate); a long-term iteration mechanism mitigates model drift and also improves the coverage of low-net-worth customers (from 13.2% to 29.8%), resolving the imbalance between outreach and conversion, and facilitating precise sales outreach. Attached Figure Description

[0019] 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 A flowchart illustrating a method for predicting the probability of a transaction from a target customer based on a regression model, as provided in this application embodiment; Figure 2 This is a schematic diagram of a device for predicting the probability of a target customer's transaction based on a regression model, provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application aims to accurately analyze the large customer base of life insurance companies. By leveraging customer characteristic information held by the life insurance companies, such as age, star rating, income, occupation, long-term insurance premiums, and short-term insurance premiums, various machine learning algorithms are used to predict the probability of closing a deal after a customer visit. This helps sales personnel identify high-potential customers. The entire system process is divided into four main steps: data preprocessing and cleaning, model training and candidate model selection, customer grouping experiments and validation, and model optimization parameter adjustment. Specifically, this application provides a method and device for predicting the probability of closing a deal with a target customer based on a regression model. The technical solution proposed in this application is described in detail below with reference to the accompanying drawings.

[0022] Figure 1 A flowchart illustrating a method for predicting the probability of a transaction with a target customer based on a regression model, as provided in this application embodiment. Figure 1 As shown, the method mainly includes the following steps: Step S1: Remove dirty data and transform unstructured data from the original customer data of the life insurance company to obtain standardized data.

[0023] Specifically, dirty data removal involves: unifying the time format to a preset standard format and removing time data that is not in the standard format; and removing ID number data that is in a preset hash value format.

[0024] The unstructured data conversion uses a tag encoding method to convert Chinese character data into a computer-recognizable numerical format, wherein a preset affirmative expression is replaced with a first number, and a preset negative expression is replaced with a second number.

[0025] Step S2: Based on customer income levels, the standardized data is divided into low-net-worth customer data sources and high-net-worth customer data sources, which are used for subsequent predictions respectively.

[0026] Specifically, the customer data source segmentation criteria are as follows: based on the customer star rating system, customers in the lower star rating range are defined as low-net-worth customers, and customers in the higher star rating range are defined as high-net-worth customers. In this target customer conversion probability prediction method, the customer segmentation criteria explicitly use the life insurance company's existing customer star rating system as the core basis. Specifically, star rating range thresholds are pre-set, clearly defining customers in the pre-set lower star rating range (such as one star and below mentioned in the document) as low-net-worth customers; and simultaneously defining customers in the pre-set higher star rating range (such as one star and above) as high-net-worth customers. This segmentation avoids interference from the key factor of customer purchasing power in subsequent predictions, allowing the two customer data sources to conduct targeted predictions separately, laying the foundation for improving prediction accuracy.

[0027] Step S3: Select multiple machine learning models as candidate models, use a dual-drive strategy to screen key features of customer transactions, and build an initial prediction model.

[0028] Specifically, the candidate models include random forest, gradient boosting decision tree, and logistic regression models. The key characteristic for selecting customer conversion rates is: if the conversion rate increases as the number of targeted customers decreases, the model is considered effective. Ultimately, the model with the highest conversion rate for the same number of targeted customers is selected as the initial prediction model. In other words, during the initial prediction model construction phase, the selected candidate models include random forest, gradient boosting decision tree, and logistic regression models, used to analyze the data sources of low-net-worth and high-net-worth customers separately. Model selection prioritizes effectiveness; that is, if the conversion rate increases as the number of targeted customers decreases, the model is considered to have the ability to effectively screen high-potential customers. Finally, from the effective models, the model with the highest conversion rate for the same number of targeted customers is selected as the initial prediction model for subsequent training.

[0029] The dual-drive strategy specifically involves: on the one hand, identifying the implicit characteristics of high-value customers through feedback from sales partners; on the other hand, conducting data-driven feature importance assessment based on feature importance evaluation methods. These methods include SHAP value analysis or regularized feature selection. The key features selected include near-term customers, customer age, customer gender, matured customers, claiming customers, renewal customers, customer annual income, universal account balance, cumulative annual payment for the year, and recent wealth management product purchases.

[0030] Step S4: Train the initial prediction model, determine the optimal parameters through regularization and cross-validation, and obtain the optimized regression prediction model.

[0031] Specifically, when training the initial prediction model, L1 regularization is used to improve prediction accuracy and avoid overfitting. This is combined with multi-fold cross-validation (such as the five-fold cross-validation mentioned in the documentation) to optimize parameters. Through repeated testing and adjustments, key optimal parameters such as the regularization coefficient, number of iterations, learning rate, and classification threshold are precisely determined. This series of optimization operations ultimately yields a more stable and accurate optimized regression prediction model.

[0032] Step S5: Perform group validation and real-world scenario validation on the optimized regression prediction model to determine the optimal transaction probability threshold, and establish a long-term iterative mechanism to continuously optimize the model.

[0033] Specifically, the long-term iterative mechanism involves: real-time tracking of customer development process data to build a dynamic evaluation system for model performance; regular collection of feedback from the sales team; and continuous optimization of the feature weight matrix and decision threshold through incremental learning and model retraining to alleviate model drift. In this regression-based method for predicting the probability of a target customer transaction, the long-term iterative mechanism is as follows: to adapt to the dynamic changes in life insurance business and effectively alleviate model drift, firstly, real-time tracking of the entire customer development process data (including core business indicators such as visit conversion rate and customer response time) is used to build a dynamic evaluation system that can monitor the model's prediction accuracy and the actual matching degree of the conversion rate in real time; secondly, regular collection of business-side feedback from the sales team, covering key aspects such as customer profile deviations and changes in the market environment; and finally, combining new data and business feedback, continuous optimization of the feature weight matrix and decision threshold through incremental learning (using new customer data to avoid full retraining to improve efficiency) and model retraining to ensure that the model always fits the actual needs of life insurance business and maintains stable and accurate prediction performance.

[0034] The group validation includes historical data validation and optimal threshold determination, with the optimal threshold determined through gradient testing. The real-world scenario validation involves continuous, pre-defined periodic tracking and observation, using the model to identify high-potential customers and calculating their conversion rates. In the validation phase of this conversion probability prediction method, the group validation comprises two core components: first, historical data validation, which excludes interference data such as those from special holiday months, selecting customer visit data from regular months as the training basis, and then using the trained model to analyze customer data from subsequent months, verifying the prediction accuracy in conjunction with the actual transaction situation of that month; second, optimal threshold determination, using gradient testing (e.g., a threshold of 10% corresponds to a conversion rate of 21%, 20% corresponds to 30.8%) to find the critical point where the conversion rate increase slows down (16% in the document) as the optimal threshold. The real-world scenario validation involves continuous, pre-defined periodic tracking and observation (e.g., three months), first using the model to identify high-potential customers and push them to the sales end, then calculating the conversion rates of these customers and unmarked customers to verify the model's effectiveness in real-world business.

[0035] In this embodiment of the application, during the data processing, the original data contains a large amount of dirty data and unquantifiable data, and it is necessary to remove dirty data (such as incorrect time format, MD5 format ID number). This invention sets YYYY-MM-DD as the standard time format and removes time formats such as YYYY.MM.DD and YYYY / MM / DD. If the ID number is 32 and is an MD5 hash value, it is removed. Unstructured data (such as Chinese characters) is converted into a computer-recognizable digital format using tag encoding, such as replacing 1 if yes and 0 if no.

[0036] During the tiered processing, customers' purchasing power has a significant impact on transactions. To avoid interference from this single factor, we divide the data into two data sources based on customer income level: "low-net-worth customers" and "high-net-worth customers," and make predictions for the two data sources separately.

[0037] Furthermore, the model building process is as follows: 1. This test selected three training models: random forest, gradient boosting decision tree, and logistic regression, and then used these three models to analyze two different data sources.

[0038] 2. Screening Key Customer Conversion Characteristics: The correlation between the number of targeted customers and the conversion rate was verified. If the conversion rate increases when the number of targeted customers decreases, the model is effective; conversely, if it increases, the model is ineffective. Based on multiple comparative experiments, logistic regression was ultimately chosen as the model for training (it achieved the optimal conversion rate with the same number of targeted customers).

[0039] 3. Key Feature Screening: There are more than 20 customer transaction-related features (star value, near-star customers, lapsed customers, customer age, customer gender, customer star rating, customer category, matured customers, claims customers, renewal customers, cumulative premiums, customer annual income, visits without a transaction, critical illness coverage, universal account balance, historical single policy of 100,000, historical single policy premium of 10,000, historical single policy premium of 30,000, three-year key customers, current year's cumulative annual premiums, number of star rating certifications, recent purchase of wealth management products, number of valid long-term insurance policies), which are difficult to cross-verify and determine manually.

[0040] To improve model predictive performance, this study employs a dual-drive key feature selection strategy: on the one hand, it integrates a manual screening mechanism to identify latent characteristics of high-value customers through feedback from business partners; on the other hand, it uses data-driven feature importance assessment (such as SHAP value analysis and L1 regularized feature selection) to retain key features that significantly explain the target variable. Ultimately, 10 key features were selected (near-term customers, customer age, customer gender, matured customers, claiming customers, renewal customers, customer annual income, universal account balance, cumulative annual payments for the current year, and recent wealth management product purchases). By dynamically adjusting feature weight allocation and feature combination strategies, while controlling the stability of the customer pool, the conversion rate of the model is improved with the same sample size, ultimately constructing an optimized model that balances predictive accuracy and business interpretability.

[0041] 4. Parameter Settings: To avoid overfitting, this invention employs L1 regularization and uses five-fold cross-validation. Testing was conducted within the range of C=0.5-1.2, with the highest model accuracy (91%) achieved at C=0.8. Therefore, the key parameters for this experiment are: regularization coefficient C=0.8, iteration count max_iter=1000, learning rate lr=0.01, and classification threshold set to 0.5.

[0042] 5. Long-term Iteration and Closed-Loop Optimization Mechanism: To ensure the model's adaptability in dynamic business scenarios, a full lifecycle iteration framework was designed: A dynamic model performance evaluation system was constructed by tracking customer development process data in real time (including metrics such as visit conversion rate and customer response timeliness); feedback from the sales team was collected regularly (such as customer profile deviations and changes in the market environment); and the feature weight matrix and decision threshold were continuously optimized through incremental learning and model retraining. This iteration mechanism achieves closed-loop optimization of "data input - model output - business feedback - parameter update," effectively mitigating model drift issues. Furthermore, the customer group experiment and verification process is as follows: 1. Grouping criteria: Customers are divided into "low-net-worth customers" and "high-net-worth customers" according to the star rating system. Customers with one star or below are low-net-worth customers, and customers with one star or above are high-net-worth customers. The signing patterns of the two groups are analyzed independently (high-star customers have a high signing rate, while low-star customers have great development potential).

[0043] 2. Validation using past data: In selecting training data, we excluded certain special months, such as months with special holidays, and selected visited customers as training clients. Then, we analyzed the customers from the months following the selection (customer characteristics were based on the data of that month). Using the analyzed data, we again selected visited customers from the selected months and combined this with the actual transaction data of that month to verify the accuracy of the prediction. If the probability threshold gradually increases, and the probability of a visit leading to a sale also gradually increases, then the model is effective. When the sale probability threshold increases to a certain value, the increase in the sale rate no longer becomes significant; this threshold represents the optimal result.

[0044] 3. Determination of optimal threshold parameters: By conducting threshold gradient tests and recording data, the following results were obtained: threshold 10%, transaction rate 21%; threshold 12%, transaction rate 25%; threshold 14%, transaction rate 28%; threshold 16%, transaction rate 30%; threshold 18%, transaction rate 30.6%; threshold 20%, transaction rate 30.8%. Therefore, 16% is the optimal threshold.

[0045] 4. Real-world verification: After determining the optimal threshold, further predictive verification experiments were conducted in real business scenarios. The specific implementation process is as follows: First, a customer value assessment model was built based on the training dataset, and this model was used to score and classify the current customer group in batches; second, a subset of high-potential customers was selected and marked based on the model output results, forming a sales priority list and pushing it to sales personnel; finally, through three consecutive months of tracking and observation, the system recorded the difference in visit conversion rates between marked and unmarked customers. The conversion rate for marked customers (selected at a 16% threshold) was 30%, while that for unmarked customers was 13%, resulting in a 131% increase in the conversion rate, and the visit coverage rate for low-net-worth customers increased from 13.2% to 29.8%.

[0046] Experimental results show that the optimized model significantly improves the sales efficiency of life insurance sales personnel. In the target customer screening stage, the model transforms the traditional visit mode into a data-driven customer segmentation visit mode, providing a data-driven decision support tool for the precise marketing of life insurance business. In actual business scenarios, this invention can significantly improve the visit-to-conversion rate of life insurance sales personnel.

[0047] The above describes a target customer transaction probability prediction system based on a regression model, as provided in this application. Based on the same inventive concept, this application also provides a target customer transaction probability prediction device based on a regression model. Figure 2 A schematic diagram of a device for predicting the probability of a target customer's transaction based on a regression model, as provided in this application embodiment, is shown below. Figure 2As shown, the device mainly includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor; wherein the memory 202 stores instructions that can be executed by the at least one processor 201, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to complete the aforementioned method for predicting the probability of a target customer's transaction based on a regression model.

[0048] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0049] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the probability of a transaction with a target customer based on a regression model, characterized in that, The method includes the following steps: Step S1: Remove dirty data and transform unstructured data from the raw customer data of the life insurance company to obtain standardized data; Step S2: Based on customer income levels, the standardized data is divided into low-net-worth customer data sources and high-net-worth customer data sources, which are used for subsequent predictions respectively; Step S3: Select multiple machine learning models as candidate models, use a dual-drive strategy to screen key features of customer transactions, and build an initial prediction model; Step S4: Train the initial prediction model, determine the optimal parameters through regularization and cross-validation, and obtain the optimized regression prediction model; Step S5: Perform group validation and real-world scenario validation on the optimized regression prediction model to determine the optimal transaction probability threshold, and establish a long-term iterative mechanism to continuously optimize the model.

2. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S1, the dirty data removal specifically involves: unifying the time format to a preset standard format and removing time data that is not in the standard format; and removing ID number data that is in the preset hash value format.

3. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S1, the unstructured data conversion uses a tag encoding method to convert Chinese character data into a computer-recognizable digital format, wherein a preset positive expression is replaced with a first number, and a preset negative expression is replaced with a second number.

4. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S2, the criteria for classifying customer data sources are as follows: based on the customer star rating system, customers in the low star rating range are defined as low-net-worth customers, and customers in the high star rating range are defined as high-net-worth customers.

5. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S3, the candidate models include random forest model, gradient boosting decision tree model and logistic regression model; the key features for screening customer transactions are: if the conversion rate increases when the number of targeted customers decreases, the model is considered effective, and finally the model with the best conversion rate under the same number of targeted customers is selected as the initial prediction model.

6. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S3, the dual-drive strategy specifically involves: on the one hand, identifying the implicit characteristics of high-value customers through feedback from sales business partners; on the other hand, conducting data-driven characteristic importance assessment based on characteristic importance assessment methods. The feature importance assessment methods include SHAP value analysis or regularized feature selection; the key features selected include near-term customers, customer age, customer gender, matured customers, claiming customers, renewal customers, customer annual income, universal account balance, cumulative annual payment in the current year, and most recently purchased wealth management products.

7. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S4, the regularization process uses L1 regularization, and the cross-validation is multi-fold cross-validation. The optimal parameters of the model are determined through regularization and cross-validation to improve model accuracy and avoid overfitting; the parameters include: regularization coefficient, number of iterations, learning rate and classification threshold.

8. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S5, the long-term iterative mechanism specifically involves: tracking customer development process data in real time to build a dynamic evaluation system for model performance, regularly collecting feedback information from the sales team, and continuously optimizing the feature weight matrix and decision threshold through incremental learning and model retraining to alleviate the model drift problem.

9. The method for predicting the probability of a transaction with a target customer based on a regression model according to claim 1, characterized in that, In step S5, the group verification includes verification of past data and determination of the optimal threshold, which is determined by gradient testing; the actual scenario verification is continuous tracking and observation for a preset period, and high-potential customers are selected based on the model and their conversion rate is calculated.

10. A device for predicting the probability of a transaction from a target customer based on a regression model, characterized in that, The device includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method for predicting the probability of a transaction with a target customer based on a regression model, as described in any one of claims 1-9.

Images