Machine learning recommendation method for precision marketing and potential customer mining

Abstract

The invention discloses a machine learning recommendation method for precision marketing and potential customer mining. The method comprises the following steps: S1, data acquisition: establishing a user tag table based on historical data and acquiring user behavior characteristics; s2, data preprocessing: preprocessing the data, including removing abnormal values, filling missing values, and converting category features into numeric features; s3, automatic box separation: performing variable analysis and automatic box separation on the preprocessed data; s4, feature processing: converting a data table obtained by binning into a numerical value form; s5, dividing a data set: dividing the processed data set into a training set and a test set; s6, stepwise regression is carried out to select features: stepwise logistic regression is used to select model key variables; s7, manual binning and modeling data set preparation are carried out; the method has the characteristics of high automation level and high decision-making efficiency.

Description

Technical Field

[0001] This invention relates to the field of machine learning and data mining technology, specifically a machine learning recommendation method for precision marketing and potential customer discovery. Background Technology

[0002] In the financial banking industry, traditional rule-based marketing methods suffer from high subjectivity and costs. To address this issue, the industry has begun using machine learning to build scorecard models to improve the objectivity of customer screening.

[0003] However, this introduces new technical challenges: First, in the model building phase, feature engineering struggles to balance efficiency and interpretability. Automated binning is efficient but has poor interpretability, while manual binning offers strong interpretability but is time-consuming.

[0004] Secondly, and more critically, lies in the model application stage. The scorecards and score threshold tables output by the model reveal an inherent contradiction: the precision and recall rates of marketing are naturally at odds. Existing application systems merely passively display this table, and business personnel still need to guesswork to select a score threshold. Therefore, it is essential to design a machine learning recommendation method for precision marketing and potential customer acquisition that can balance feature engineering efficiency and interpretability while providing intelligent multi-objective decision support. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning recommendation method for precision marketing and potential customer mining, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a machine learning recommendation method for precision marketing and potential customer acquisition, comprising the following steps: S1. Data Acquisition: Establish a user tag table based on historical data and acquire user behavior characteristics; S2. Data preprocessing: Preprocess the data, including removing outliers, filling in missing values, and converting categorical features into numerical features; S3, Automatic Binning: Performs variable analysis and automatic binning on preprocessed data; S4. Feature processing: Convert the data table obtained from binning into numerical form; S5. Split the dataset: Divide the processed dataset into a training set and a test set; S6. Stepwise regression for feature selection: Use stepwise logistic regression to select key variables for the model; S7. Manual binning and modeling dataset preparation: Repeat steps S3 to S5 for the key feature variable combinations selected in step S6, wherein the automatic binning in step S3 is replaced by manual binning. S8. Model Training and Parameter Tuning: Input the training set from step S7 into the logistic regression model for training; S9. Model Parameter Export and Prediction Probability Calculation: Export model parameters and calculate prediction probabilities; S10. Model Evaluation: Use evaluation metrics to evaluate model performance; S11. Generate a scoring card format for the model: Convert the logistic regression model into a scoring card; S12. Generate rating threshold table and score distribution chart: Rate the users in the dataset and generate rating threshold table and score distribution chart; S13. Model Deployment and Application: Use the scoring card and scoring threshold table to assist business personnel in user selection.

[0007] According to the above technical solution, the step S1 of acquiring data further includes: Based on the pre-selection by the business department, multiple marketing scenarios are defined, with each scenario corresponding to feature combination X and category label y; The marketing scenarios include: deposits not used for initial purchase, deposits not used for repeat purchase, and deposit losses not recovered.

[0008] According to the above technical solution, the S3 automatic box sorting adopts the KS automatic box sorting method; the S7 manual box sorting adopts the badrate manual box sorting method.

[0009] According to the above technical solution, the S4 feature processing includes: S401. WOE Encoding Conversion: Calculate the WOE value for each group of each feature and map the grouped data to a WOE value data table; S402. Normalization: Scales the WOE data to the range [0,1].

[0010] According to the above technical solution, the stepwise regression in S6 selects features based on AUC as the model performance basis, and sets a termination condition for the AUC growth rate.

[0011] According to the above technical solution, the S8 model training and parameter tuning uses L2 regularization, and grid search is used to adjust the two hyperparameters: regularization strength C and weight distribution classweight.

[0012] According to the above technical solution, S9 further includes S902 model parameter inverse normalization: The weights and thresholds of the exported WOE data are denormalized; Obtain the weight w and threshold b for the original WOE data; Substitute into the probability expression Calculate the predicted probability.

[0013] According to the above technical solution, the conversion formula for generating the scorecard in S11 is as follows: And by setting , , The factor and offset are indirectly calculated from the four parameters.

[0014] According to the above technical solution, the deployment and application of the S13 model includes: S1301. Scorecard Deployment: Convert the scorecard into a code script and add the scores of the three scenarios as tags to the personalized user tag table; S1302. Deployment of the scoring threshold table: Push the content of the scoring threshold table to the business database; S1303. Model Application: Business personnel access user rating and rating threshold tables in the customer analysis management system to display the estimated reach effect.

[0015] According to the above technical solution, the S1303 model application further includes a reverse optimization recommendation method based on multi-dimensional business constraints, which includes: Obtain cost constraints, benefit constraints, and coverage constraints input by business personnel; The system iterates through the scoring threshold table and, in conjunction with the scoring data from day T-1, checks whether each threshold simultaneously satisfies all constraints. When a feasible solution exists, an optimal solution is recommended. When no feasible solution is found, the system identifies and reports conflicts and automatically calculates and recommends Pareto optimal options, including benefit-first and coverage-first strategies.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are: the present invention, (1) After using KS automatic binning, badrate manual binning is used to supplement key features. This combines the simplicity of automatic binning with the strong interpretability of manual binning, and obtains binning basis with better controllability and stronger interpretability while ensuring efficiency. (2) By performing WOE encoding conversion and normalization on the features, the variable type is numericalized and the variable contribution is monotonicized, which is conducive to the logistic regression model to obtain better performance; (3) By using stepwise logistic regression to select features, we can select as few features as possible while maintaining the modeling effect, making the model less prone to overfitting and shortening the training time. (4) By converting the complex logistic regression model into an intuitive scorecard and score threshold table, the model results are easy to understand and deploy, and are convenient for business personnel to refer to and call upon in practical applications; (5) By dynamically combining the static threshold table with the real-time scoring data of day T-1, business personnel are provided with real-time decision support based on the latest data; (6) By upgrading business decision-making from passively querying thresholds to actively inputting targets, the system automatically solves the optimal strategy. In particular, when business objectives conflict with each other, it can recommend Pareto optimal options, which significantly improves the automation level and decision-making efficiency of precision marketing. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a technical architecture diagram of the method of the present invention; Figure 2 This is a schematic diagram illustrating the specific execution flow of the method of the present invention; Figure 3 This is a schematic diagram illustrating the specific steps of data acquisition in this invention; Figure 4 This is a schematic diagram illustrating the specific steps of data preprocessing in this invention; Figure 5 This is a schematic diagram illustrating the specific steps of the feature processing of the present invention; Figure 6 This is a schematic diagram illustrating the specific steps of the stepwise regression feature selection method of the present invention; Figure 7 This represents a gradual return to the initial purchase scenario of this invention; Figure 8 This is a schematic diagram illustrating the specific steps of model parameter derivation and prediction probability calculation in this invention; Figure 9 This is a schematic diagram illustrating the specific steps of the scoring card format of the generative model of the present invention; Figure 10 This is a schematic diagram of the scoring card for the first purchase scenario of this invention; Figure 11 This is a schematic diagram illustrating the specific steps of generating the scoring threshold table and score distribution chart according to the present invention; Figure 12 This is an intentional representation of the scoring threshold in the first-purchase scenario of this invention; Figure 13 This is a score distribution chart for the first purchase scenario of this invention; Figure 14 This is a schematic diagram illustrating the specific steps of model deployment and application of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The technical architecture of a machine learning recommendation method for precision marketing and potential customer mining, as described in this invention, mainly includes three parts: model training, model deployment, and model application. This invention provides the following technical solution: Figure 2 As shown, a machine learning recommendation method for precision marketing and potential customer acquisition includes: S1. Data Acquisition: Based on historical data, establish a personalized user tag table and acquire user behavior characteristics.

[0020] Specifically, features highly correlated with mobile banking deposit and purchase behavior are identified, including login, real-name authentication, account opening, participation in activities, deposit, wealth management, and loan information. These features are used to construct a dataset for machine learning model training, parameter tuning, and prediction.

[0021] like Figure 3 As shown, three marketing scenarios are illustrated. The business department pre-selects feature combinations X and classification labels y with high relevance to each scenario: (X1, y1), (X2, y2), and (X3, y3). It is agreed that for y1, y2, and y3, bad samples are recorded as 1, and good samples as 0. y1 represents the deposit that has not been purchased before. If the user has not purchased a deposit, y1 = 1; if the user has purchased a deposit, y1 = 0. y2 represents the deposit not being repurchased. If the user has not completed the deposit repurchase, y2 = 1; if the user has completed the deposit repurchase, y2 = 0. y3 represents the deposit loss that has not been recovered. If a user is lost (deposit balance = 0 for X consecutive days) and has not been recovered to date, y3 = 1. If a user is lost but has been recovered, y3 = 0.

[0022] After determining the characteristics, retrieve the specified user data from the personalized user tag table. Specifically: The first-time purchase scenario selects all mobile banking users; Repeat purchases are restricted to users who have already made their first mobile banking purchase. The churn recall scenario selects cases where all users have churned (those who have made their first purchase and have had 0 days of deposit). Records (30 days) are used to extract feature data for each record before and during the churn period. If the same user has multiple churn and recall records, they are considered as multiple samples.

[0023] S2. Data Preprocessing: Preprocessing the data, including removing outliers, filling in missing values, and converting categorical features to numerical features.

[0024] like Figure 4 As shown, optionally, the specific steps of data preprocessing in step S2 are as follows: S201. Outlier Handling: Outlier handling is performed based on the actual situation of the data, which typically includes: deleting outlier samples, replacing outliers with other values, and transforming outliers.

[0025] S202. Missing value imputation: Imput missing values ​​according to the actual situation of the data, such as deleting missing value samples, linear interpolation, median, mean, mode, setting to 0, setting to -1, etc.

[0026] S203. Transforming categorical features into numerical features: This typically includes dictionary mapping, integer encoding, and one-hot encoding.

[0027] S3. Automatic binning: Performs variable analysis and binning on preprocessed data, and selects variables that are highly correlated with user quality as input features for modeling.

[0028] In this embodiment of the invention, variable binning discretizes the originally continuous feature values ​​into multiple bins. Binning methods include automatic binning (equal-distance binning / equal-frequency binning / chi-square binning / KS binning / decision tree binning, etc.) and manual binning. After binning, the percentage of bad samples within each bin is obtained by statistically analyzing the samples. If each bin has a certain characteristic If there is a clear correlation between the feature value intervals corresponding to each bin, then the feature can be included in the model.

[0029] Automated sorting can fix the process and simplify execution once the method is defined; however, its sorting criteria are not intuitive enough, have weak interpretability, and are easily affected by extreme values. Manual sorting, on the other hand, has the advantages of good controllability and strong interpretability of sorting criteria; however, each feature is adjusted step by step, making the operation relatively time-consuming.

[0030] After testing, KS automatic box sorting and Manual binning, which yields the same optimal feature combinations after subsequent selection, results in similar model performance. This application combines the advantages of both methods. In this step, KS automatic binning is used first for rapid feature selection, followed by manual binning of the selected features in subsequent steps to obtain binning criteria that are more controllable and interpretable.

[0031] Specifically, the KS automatic binning method works as follows: Each bin is divided in half at its maximum KS value (based on the ratio of good to bad samples) until the number of bins reaches a set value, or the number of samples in each bin falls below a set value. This results in a group data table in interval format. The formula for calculating the maximum KS value is: S4. Feature processing: Convert the data table in text interval format obtained from binning into numerical format; For example, in this embodiment of the invention, WOE evidence weight mapping is used: based on the proportion of good and bad samples, the value range of features is converted into WOE values. WOE mapping has the characteristics of numericalizing variable types, monotonicizing variable contributions, and rationalizing variable contributions.

[0032] Converting the data from interval format to WOE values ​​and then normalizing the WOE data helps to obtain better model performance. like Figure 5 As shown, the specific steps of the feature processing in step S4 are as follows: S401. WOE Encoding Conversion: Calculate the WOE value for each group of each feature, generate a WOE value mapping table for the groups, and then map the interval-formatted grouped data into a WOE value data table. For the i-th group of feature X, the WOE value calculation formula is as follows: in, This represents the total number of bad samples. The total number of good samples, Let X be the number of bad samples in the i-th group of feature X. Let X be the number of good samples in the i-th group of feature X.

[0033] S402. Normalization: This is also known as minimum-maximum scaling, which scales WOE data to the range [0,1]. Normalization formula: in, They are respectively A row vector consisting of the maximum and minimum values ​​in each column; This means subtracting X from each row. ; express Each line is respectively with Divide corresponding elements.

[0034] S5. Split the dataset: Divide the processed dataset into a training set and a test set for the model data in an 8:2 ratio to maintain a balance between positive and negative samples.

[0035] S6. Stepwise Regression for Feature Selection: Stepwise logistic regression is used to select key variables for the model. While maintaining the modeling effect, as few features as possible are selected to prevent the model from overfitting and to shorten the training time.

[0036] like Figure 6 As shown, the specific steps for selecting features through stepwise regression in step S6 are as follows: S601. Select the optimization direction: Use AUC as the basis for model performance.

[0037] S602. Set termination conditions: ; where n is the number of rounds of stepwise regression.

[0038] S603. Perform the stepwise regression process: Iterate through all variables, model each variable against the target, and select the variable with the best model result as the first round of selection variable; Based on the first round of variable selection, a second variable is added: iterate through the remaining variables, and select the variable that best results in the model as the second round of variable selection; Based on the second round, add a third variable; Adding variables stops when new variables no longer significantly contribute to the fitting results (meeting the termination condition), and the previously added variables are used as the modeling feature set. For example, such as Figure 7 The flowchart shown is a regression diagram for the first-time purchase scenario.

[0039] S7. Manual binning and modeling dataset preparation: Repeat steps S3-S5 for the key feature variable combinations selected in step S6, where step S3 is changed from KS automatic binning to... Manually bin the data to obtain training and test sets.

[0040] Manual box sorting method: First observe Based on the changing trend of eigenvalues, several bins were initially divided, and then gradually adjusted and merged. Bins with similar values.

[0041] S8. Model Training and Parameter Tuning: Logistic regression is a classification algorithm based on regression analysis, and its formula is as follows: Logistic regression adds a sigmoid function to linear regression, mapping the data to the interval (0, 1) and dividing it into two categories based on a threshold. Data greater than or equal to the threshold is grouped into one category, and data less than the threshold is grouped into another. The expression for the sigmoid function is as follows: The training set from step S7 is input into the logistic regression model for training, L2 regularization is applied, and two important hyperparameters are tuned using grid search: 1. Regularization strength C: A lower value indicates strong regularization, which suppresses model complexity and may lead to underfitting. Conversely, a higher value may lead to overfitting. 2. Weight Distribution Setting it to 'balanced' makes the positive and negative samples more balanced, while setting it to None has no effect.

[0042] S9. Model parameter export and prediction probability calculation: After model training is complete, the weights (coefficients) and thresholds for each feature are derived. Since the WOE data was normalized before training, the directly extracted weights and thresholds are based on the normalized WOE data. Denormalization is then performed to obtain the weights w and thresholds b for the original WOE data x. These are then substituted into the model's probability expression: This allows us to calculate the probability that the model predicts a bad sample.

[0043] like Figure 8 As shown, the specific steps for deriving model parameters and calculating prediction probabilities in step S9 are as follows: S901. Model Parameter Export: Using the logreg.coef_ and logreg.intercept_ methods in the sklearn library, the weights (coefficients) and thresholds of each feature of the model can be obtained.

[0044] S902. Inverse Normalization of Model Parameters: Calculation Method: Due to the normalization formula: ; Therefore, the linear part of the model: ; Therefore, the weights of the original WOE data Threshold of raw WOE data .

[0045] S903. Prediction Probability Calculation: Substitute the weights w and threshold b for the original WOE data x into the model's probability expression: This allows us to calculate the probability that the model predicts a bad sample.

[0046] S10, Model Evaluation This invention uses evaluation metrics such as the area under the ROC curve (AUC), precision (P), and recall (R) to evaluate the performance of the model.

[0047] The relevant definitions are as follows: In the confusion matrix, T and F represent whether the prediction is correct, and P and N represent positive samples (1, bad users) and negative samples (0, good users).

[0048] AUC: The area under the ROC curve (FPR-TPR curve), representing the model's ability to distinguish between positive and negative samples, where FPR=FP / (TN+FP) and TPR=TP / (FN+TP).

[0049] Good user precision: TN / (TN+FN), which represents the percentage of samples where the predicted value is a good user and the actual value is a good user, based on the prediction results. Good user recall rate: TN / (TN+FP), which is the percentage of samples with predicted values ​​of good users among samples with true values ​​of good users. Bad User Precision: TP / (FP+TP), which represents the percentage of samples where the predicted value is a bad user and the actual value is a bad user, based on the prediction result. Bad user recall rate: TP / (FN+TP), which is the percentage of samples whose predicted value is a bad user among samples whose true value is a bad user. In this invention, precision affects the success rate of marketing, while recall affects the coverage rate. Since business needs prioritize covering as many target users as possible, the performance evaluation and optimization objective of the model in this invention is set as a good user recall rate.

[0050] S11, Scorecard format for generative models Creating a scorecard is essentially converting the model into a score. After training, the linear component of the logistic regression model is extracted and linearly scaled to create a scorecard table. When using it, a correspondence is established between feature value, feature group, and feature score for each feature of the sample. The scores of each feature are summed, and then added to the base score to obtain the total score.

[0051] Conversion formula: Where Score is the total score, offset and factor are manually set parameters, BaseScore is the base score, FeatureScore is the feature score of a single feature and a single group, w is the model coefficient of a single feature, x is the WOE value of a single feature and a single group, and b is the model threshold.

[0052] like Figure 9 As shown, the specific steps for generating the scorecard form of the model in step S11 are as follows: S1101. Setting Indirect Parameters: Factor and offset are generally not set directly, but are set through... , , The four parameters are indirectly calculated, and their meanings are as follows (odds: the ratio of the probability of a bad user to the probability of a good user): , :odds is At that time, the rating was ; , : every decrease in odds When the score is doubled, the score increases. .

[0053] The default value of the parameter is: = 600, = 50, = 20, = 2.

[0054] This means that when odds = 50:1, the score is 600; for every halving of odds (e.g., from 50:1 to 25:1), the score increases by 20.

[0055] S1102. Calculate parameters factor and offset: After setting the four parameters according to business needs (this invention uses the default values ​​in S1101), determine factor and offset according to the following formula: S1103. Calculate the base score (BaseScore) and the feature score (FeatureScore) for each feature and each group: S1104. Creating a Scorecard: Based on BaseScore and FeatureScore, create a scorecard format for the model. For example, ... Figure 10 As shown, the scoring card for the first purchase scenario consists of a BaseScore and a FeatureScore for each feature in each group.

[0056] S12. Generate a scoring threshold table and score distribution chart: Based on the scorecard obtained in step S11, each user in the dataset can be scored. After scoring, by statistically analyzing the number of users, good users, and bad users within each score interval (interval length is usually set to 10, such as 700~710 points), a distribution chart of good and bad users in each score interval (score distribution chart) can be drawn, and user metrics (scoring threshold table) under different score thresholds (such as greater than or equal to 700 points) can be calculated for business reference. This invention selects the following metrics to form the scoring threshold table: 1. Selected user percentage, 2. Selected percentage of good users (good user precision), 3. Selected percentage of good users (good user recall).

[0057] The scoring threshold table reflects the performance of past data (modeling dataset) and serves as a prediction and reference for future data. As the selected score threshold increases, user quality gradually improves, increasing the percentage of good users among the selected users (good user precision); good users with lower scores are filtered out, decreasing the percentage of selected good users (good user recall). In practical applications, the selected score threshold can be flexibly adjusted according to the business's needs for precision and recall. This is derived from the model's probability expression. Adjusting the score threshold is equivalent to adjusting the probability P threshold of the model (default value P=0.5, corresponding to Score=offset=712.877).

[0058] like Figure 11 As shown, the specific steps for generating the scoring threshold table and score distribution map in step S12 are as follows: S1201. Calculate the score: Using the scorecard obtained in step S11, score each feature of each sample (including the training set and the test set), sum the scores of each feature, and add the base score to obtain the total score.

[0059] S1202. Interval Statistics: Divide the sample into score intervals (each interval length is set to 10), and count the number of good and bad samples in each score interval. In addition, other label statistics can be added according to business needs.

[0060] S1203. Determine the scoring threshold table indicators: To reflect the model performance under different score thresholds, this invention selects the following indicators: 'proportion of selected users', 'proportion of good users among selected users (precision)', and 'proportion of good users selected (recall)'. In addition, other indicators can be added according to business needs.

[0061] S1204. Generate a scoring threshold table: Calculate and process the statistical data in step S1202 to obtain the index values ​​in step S1203 under each score threshold.

[0062] S1205. Generate score distribution chart: Calculate and process the statistical data in step S1202 to obtain the distribution chart of good and bad samples in each score interval.

[0063] For example, such as Figure 13 As shown, the score distribution chart for the first purchase scenario reflects the distribution of users who have made a first purchase / those who have not made a first purchase in each score range. The horizontal axis value corresponds to the range from the current score to the next score, such as '480' corresponding to the score range [480, 490].

[0064] S13, Model Deployment and Application In this embodiment of the invention, the scoring cards and scoring threshold tables output by the model are used to assist business personnel in selecting users and to help achieve precise marketing reach.

[0065] When deploying the scorecard, it is converted into a code script that calculates scores based on user characteristics. The three scores for three scenarios—first deposit purchase, repeat deposit purchase, and deposit loss recall—are used as three tags and added to the existing personalized user tag table. The script is then deployed on the WS platform to execute scheduling tasks and calculate scores for all mobile banking users daily.

[0066] When the scoring threshold table is deployed, it is pushed to the business database for data entry, and updated along with the model.

[0067] The model is applied as follows: Business personnel, building upon the existing customer segmentation in the customer analysis and management system, use the marketing module for further filtering. The marketing module retrieves user ratings from the personalized user tag table (T-1 day) and the latest rating threshold table, displaying the number of selected users, the percentage of selected users, and the estimated reach (precision, recall) for each score threshold within the already selected customer segment, highlighting the recommended thresholds. The threshold, precision, and recall columns, along with whether a threshold is recommended, are directly taken from the rating threshold table. The number of selected users and the percentage of selected users are derived from a statistical comparison of user ratings and thresholds in the personalized user tag table.

[0068] like Figure 14 As shown, the specific steps for model deployment and application in step S13 are as follows: S1301. Scorecard Deployment: Convert the scorecard into a code script that calculates scores based on user characteristics. Use the three scores for three scenarios—first deposit purchase, repeat deposit purchase, and deposit loss recovery—as three tags, add them to the existing personalized user tag table, and deploy them on the WS platform to execute scheduling tasks, calculating scores for all mobile banking users daily.

[0069] S1302. Deployment of scoring threshold table: Push the scoring threshold table content to the business database and update it along with the model update.

[0070] S1303. Model Application: Business personnel, building upon the existing customer segmentation in the customer analysis management system, use the marketing module for further filtering. The marketing module retrieves the user ratings from the personalized user tag table (T-1 day) and the latest rating threshold table, displaying the number of selected users, the percentage of selected users, and the estimated reach (precision, recall) for each score threshold within the selected customer segment. It also highlights the recommended selection thresholds. The threshold, precision, recall, and recommendation columns are directly taken from the rating threshold table, while the number of selected users and the percentage of selected users are derived from a statistical comparison of user ratings and thresholds.

[0071] As a preferred implementation, after implementing the basic display functions described above, the S1303 model application step may further include a reverse optimization recommendation method based on multi-dimensional business constraints. This method automatically solves for the optimal score threshold by taking the business objective as input, specifically including: Obtaining Multi-Dimensional Business Constraint Input: The marketing module further provides a decision configuration interface for receiving a set of conflicting business constraints input by business personnel, rather than requiring them to directly select score thresholds. The business constraints include at least: Cost constraint: The maximum number of marketing contacts or the maximum percentage of selected users that business personnel can accept.

[0072] Benefit constraint: The minimum marketing conversion rate P required by the sales staff, which corresponds to the percentage of good users (accuracy) among the selected users in the rating threshold table generated in S12.

[0073] Coverage constraint: The minimum marketing coverage R desired by business personnel, which corresponds to the "Percentage of Selected Good Users (Recall)" in the scoring threshold table generated in S12. Constraint Solving and Conflict Detection: After receiving the business constraints, the system automatically executes the following constraint solving process: The system iterates through all score threshold rows in the scoring threshold table generated by S12.

[0074] For each score threshold, perform: Static query: Retrieve the estimated accuracy corresponding to the given threshold from the scoring threshold table. and estimated recall rate .

[0075] Dynamic statistics: Based on the T-1 day rating data of all users deployed on S1301, the system calculates in real time the number of actual selected users whose ratings are greater than or equal to the threshold within the selected customer groups (e.g., the "non-first-time purchase" group). .

[0076] Constraint Comparison: Check whether the threshold simultaneously meets the three constraints input by the business personnel. and and .

[0077] Based on the results of constraint comparison, the system outputs the optimal decision to business personnel: When a feasible solution exists, meaning the system finds one or more score thresholds that simultaneously satisfy all constraints, the system will calculate and recommend an optimal solution (e.g., one that satisfies both P and R constraints, making...). The smallest threshold), and highlight the recommended threshold in the S1303 interface.

[0078] When there is no feasible solution, the system finds that no single score threshold can satisfy all constraints simultaneously (this is common in business because precision and recall are inherently conflicting).

[0079] At this point, the system will proactively identify and report the conflict (e.g., "Your settings of 'precision >= 60%' and 'recall >= 70%' cannot be met simultaneously").

[0080] The system automatically calculates and recommends two Pareto optimal options (non-dominated solutions) for business personnel, for example: Strategy A (Efficiency First): Recommended "When satisfying..." Under the premise of [the above conditions], the highest recall rate that can be achieved The corresponding score threshold.

[0081] Strategy B (Coverage Priority): Recommends "when satisfying..." Under the premise of [the above conditions], the highest accuracy that can be achieved The corresponding score threshold.

[0082] Sales personnel can determine the final customer segmentation plan with one click based on the system's recommended strategy A or B.

[0083] This reverse optimization recommendation method dynamically and intelligently combines the static threshold table of S12 with the real-time scoring data of S1301, solving the technical problem of business personnel blindly selecting thresholds when faced with multiple conflicting objectives. It upgrades the model application from passive query to proactive decision support, significantly improving the automation level and decision-making efficiency of precision marketing.

[0084] S14, Model Update and Optimization For example, in this embodiment of the invention, the marketing segmentation model updates its training data and remodels approximately every 3 months. Business personnel conduct marketing outreach on the user list output after the filtering in step S13, and collect the data 2 weeks later to test the model's effectiveness on the new data, so as to further optimize the model.

[0085] The specific steps for model updating and optimization in step S14 are as follows: The marketing model of this invention updates its training data and remodels approximately every 3 months. Business personnel conduct marketing outreach to the user list output after the filtering in step S13, and collect the data 2 weeks later to test the model's effectiveness on new data, in order to further optimize the model.

[0086] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A machine learning recommendation method for precision marketing and potential customer acquisition, characterized in that, Includes the following steps: S1. Data Acquisition: Establish a user tag table based on historical data and acquire user behavior characteristics; S2. Data preprocessing: Preprocess the data, including removing outliers, filling in missing values, and converting categorical features into numerical features; S3, Automatic Binning: Performs variable analysis and automatic binning on preprocessed data; S4. Feature processing: Convert the data table obtained from binning into numerical form; S5. Split the dataset: Divide the processed dataset into a training set and a test set; S6. Stepwise regression for feature selection: Use stepwise logistic regression to select key variables for the model; S7. Manual binning and modeling dataset preparation: Repeat steps S3 to S5 for the key feature variable combinations selected in step S6, wherein the automatic binning in step S3 is replaced by manual binning. S8. Model Training and Parameter Tuning: Input the training set from step S7 into the logistic regression model for training; S9. Model Parameter Export and Prediction Probability Calculation: Export model parameters and calculate prediction probabilities; S10. Model Evaluation: Use evaluation metrics to evaluate model performance; S11. Generate a scoring card format for the model: Convert the logistic regression model into a scoring card; S12. Generate rating threshold table and score distribution chart: Rate the users in the dataset and generate rating threshold table and score distribution chart; S13. Model Deployment and Application: Use the scoring card and scoring threshold table to assist business personnel in user selection.

2. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The S1 data acquisition further includes: Based on the pre-selection by the business department, multiple marketing scenarios are defined, with each scenario corresponding to feature combination X and category label y; The marketing scenarios include: deposits not used for initial purchase, deposits not used for repeat purchase, and deposit losses not recovered.

3. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The S3 automatic box sorting adopts the KS automatic box sorting method; the S7 manual box sorting adopts the badrate manual box sorting method.

4. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The S4 feature processing includes: S401.WOE Encoding Conversion: Calculate the WOE value for each group of each feature and map the grouped data to a WOE value data table; S402. Normalization: Scales the WOE data to the range [0,1].

5. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The S6 stepwise regression selects features based on AUC as the model performance criterion and sets a termination condition for the AUC growth rate.

6. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The S8 model was trained and tuned using L2 regularization, and grid search was used to adjust the two hyperparameters: regularization strength C and weight distribution classweight.

7. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: S9 further includes S902 model parameter inverse normalization: The weights and thresholds of the exported WOE data are denormalized; Obtain the weight w and threshold b for the original WOE data; Substitute into the probability expression Calculate the predicted probability.

8. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The conversion formula for generating the scorecard in S11 is as follows: And by setting , , The factor and offset are indirectly calculated from the four parameters.

9. The machine learning recommendation method for precision marketing and potential customer mining according to claim 1, characterized in that: The deployment and application of the S13 model include: S1301. Scorecard Deployment: Convert the scorecard into a code script and add the scores of the three scenarios as tags to the personalized user tag table; S1302. Deployment of scoring threshold table: Push the scoring threshold table content to the business database; S1303. Model Application: Business personnel access user rating and rating threshold tables in the customer analysis management system to display the estimated reach effect.

10. The machine learning recommendation method for precision marketing and potential customer mining according to claim 9, characterized in that: The S1303 model application further includes a reverse optimization recommendation method based on multi-dimensional business constraints, which includes: Obtain cost constraints, benefit constraints, and coverage constraints input by business personnel; The system iterates through the scoring threshold table and, in conjunction with the scoring data from day T-1, checks whether each threshold simultaneously satisfies all constraints. When a feasible solution exists, an optimal solution is recommended. When no feasible solution is found, the system identifies and reports conflicts and automatically calculates and recommends Pareto optimal options, including benefit-first and coverage-first strategies.

Images