Excitation method and system based on user category purchase prediction model
By constructing a general user category purchase prediction framework and utilizing BG/NBD and lightGBM models, the problem of high-cost customized feature modeling in omnichannel e-commerce platforms is solved, enabling flexible marketing strategy formulation and high prediction accuracy.
Patent Information
- Application Number
- CN202511350458.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2026-02-06
AI Technical Summary
On omnichannel e-commerce platforms, customizing and modeling features for different business areas individually leads to high computational and human resource costs, making it difficult to flexibly respond to diverse marketing scenarios.
A general user category purchase prediction framework is constructed. Based on rich customer characteristics and parameter configurations, the BG/NBD model and lightGBM classifier are used to build a prediction model for member login retention and order probability. Combined with bucket validation and optimization modules, a differentiated coupon distribution strategy is formulated.
It improved operational efficiency, flexibly responded to category order prediction tasks in different scenarios, reduced computing and labor costs, and improved prediction accuracy and marketing effectiveness.
Smart Images
Figure CN121481606A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of operation promotion, and in particular to an incentive method and system based on a user category purchase prediction model. BACKGROUND
[0002] In today's digital age, marketing algorithms play a crucial role. These algorithms use big data and machine learning techniques to analyze user behavior and historical data to predict user purchase intentions. Predicting the probability of placing an order in a category is one of the important tasks. Based on the probability of placing an order in a category, users are stratified, and different coupons are issued to users in different strata, which is a common and effective personalized marketing strategy.
[0003] Compared with CDP (Customer Data Platform) tag circle person marketing, the coupon based on machine learning prediction of order probability is more personalized and refined. Machine learning models can learn the key factors driving each user's future order based on higher-dimensional user features. Machine learning models can be customized and adjusted according to different business needs and situations, so as to better adapt to different marketing scenarios. In contrast, CDP circles may be limited by platform and technology, with lower flexibility.
[0004] For a full-channel e-commerce platform, it operates online B2C, department store and supermarket O2O businesses. When setting coupons, there will be coupons suitable for multiple formats, coupons suitable for multiple categories, or coupons with high thresholds that only hope to find users who can bring high consumption. Complex and diverse business scenarios bring many challenges. If features are customized and modeled separately for each business range, there will be high computing and manpower costs. Therefore, we conceived a general user category purchase prediction framework based on rich user features, configured business ranges and prediction targets in the form of parameters, and flexibly coped with different scenarios of category order prediction tasks. SUMMARY
[0005] The purpose of the present application is to cope with the challenges brought by complex and diverse business scenarios. If features are customized and modeled separately for each business range, there will be high computing and manpower costs. Therefore, we conceived a general user category purchase prediction framework based on rich customer features, configured business ranges and prediction targets in the form of parameters, and flexibly coped with different scenarios of category order prediction tasks, providing an incentive method and system based on a user category purchase prediction model to solve the problems raised in the background art.
[0006] To achieve the above-mentioned application purposes, one aspect of the present application provides an incentive method based on a user category purchase prediction model, comprising the following steps:
[0007] Step S1, determine the business coupon marketing target, the marketing target includes format, marketing category, member range, amount range and activity duration;
[0008] Step S2, based on the BG / NBD (Beta Geometric / Negative Binomial Distribution) model, a member login retention estimation model is constructed, trained and the member login retention probability is predicted;
[0009] Step S3, based on the lightGBM classifier model, a member target category order probability prediction model is constructed and trained;
[0010] Step S4, the member order probability is predicted, and the actual order recall rate and precision rate are verified by bucketing;
[0011] Step S5, according to the order probability, customer CLV (Customer Lifetime Value), login retention probability, a coupon distribution strategy is formulated;
[0012] Step S6, according to the formulated coupon distribution strategy, a marketing activity is carried out, and the member login retention estimation model and the member target category order probability prediction model are optimized and adjusted according to the activity result.
[0013] Further, in step S2, the member churn rate p is subject to Beta distribution with shape parameters a and b, and the parameters a and b of the Beta distribution are learned by fitting the characteristic data recency / frequency / T of the BHL member login through MLE (Maximum Likelihood Estimation).
[0014] Further, in step S3, the positive samples are up-sampled by SMOTE algorithm, so that the positive and negative sample ratio is controlled at 1:8.
[0015] Further, in step S3, the method for determining the characteristics of the model is:
[0016] Step S301, the model characteristics include member characteristics and the combination of characteristics in the target format category;
[0017] Step S302, the characteristics of different prediction targets are determined;
[0018] Step S303, the characteristics are cleaned according to the requirements of the lightGBM model;
[0019] Step S304, the characteristics are selected based on SelectKBest of sklearn.
[0020] Furthermore, the member target category order probability prediction model calculates the characteristics of members and member categories on a daily basis based on the data tracking details of the past year, updates the characteristics, and inserts them into the daily partition.
[0021] Furthermore, the member target category order probability prediction model predicts all member characteristics on a future day by configuring time parameters.
[0022] Furthermore, in step S3, the model is tested offline and trained online on the full dataset, respectively:
[0023] Offline testing involves dividing the training data into two parts, with 80% used for training and 20% for testing. The optimal hyperparameters of the model are obtained through training.
[0024] The full-data online training of the model involves training the lightGBM classifier daily using the full dataset.
[0025] Furthermore, in step S4, the bucket verification first uses the model parameters selected from the offline test to predict the probability of a member placing an order for a product category in the next N days. The members are then sorted and bucketed according to their order probability, and the order probability of each bucket is verified.
[0026] Furthermore, the model was refined by analyzing key features for different product categories using tools, including the FeatureImportance function and SHAP value built into lightGBM.
[0027] Another aspect of the present invention provides an incentive method based on a user category purchase prediction model, comprising a marketing objective module, a retention prediction module, an order prediction module, a bucket validation module, a delivery module, and an optimization module, wherein:
[0028] The marketing objectives module is used to determine the marketing objectives for business coupons. The marketing objectives include business type, marketing category, membership scope, amount range, and activity duration.
[0029] The retention prediction module is used to build a member login retention prediction model based on the BG / NBD model, train it, and predict the member login retention probability.
[0030] The order prediction module is used to build and train a member target category order probability prediction model based on the lightGBM classifier model;
[0031] The bucket verification module is used to predict the probability of a member placing an order and to perform bucket verification on the recall and precision of the actual orders.
[0032] The campaign delivery module is used to formulate coupon delivery strategies based on order placement probability, customer CLV, and login retention probability.
[0033] The optimization module is used for carrying out a marketing activity according to a formulated coupon distribution strategy, and optimizing and adjusting a member login retention estimation model and a member target category order probability prediction model according to an activity result.
[0034] Compared with the prior art, the system and method have the following advantages:
[0035] The present application supports various customer features and customer feature combinations in a target format category, and configures the business scope and prediction target in the form of parameters to improve the operation efficiency in a configured manner and flexibly cope with different scene category order prediction tasks. BRIEF DESCRIPTION OF DRAWINGS
[0036] Figure 1 A flowchart of an incentive method based on a user category purchase prediction model. DETAILED DESCRIPTION
[0037] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.
[0038] As Figure 1 shown is a method flowchart of the present application, and the embodiment of the present application provides an incentive method based on a user category purchase prediction model, and the specific steps are as follows:
[0039] Step S1, determining a business coupon marketing target, wherein the marketing target includes a format, a marketing category, a member range, an amount range and an activity duration.
[0040] The format is a marketing channel, including a main station B2C, a supermarket fast home and a department store cloud store. The category range of the marketing category can be single-selected, multi-selected, a first-level or second-level category commodity, or some second-level category commodities are excluded. The member range is a member who has placed an order in a target category in the past year or has viewed a target category commodity detail page in the past 60 days. The amount range needs to limit the members who have consumed in the amount range of the category as positive samples.
[0041] The activity duration is a time range for determining a prediction target. Model training and prediction need to have a time range defined. For example, if the time range is set to 10 days, the member features in the past 10 days are used as X in the training phase, and whether there is an order in the target range in the past 10 days is used as y to train the model. In the prediction phase, the latest member features X of today are input into the model to obtain the member category order probability in the future 10 days.
[0042] Step S2, a member login retention prediction model is constructed based on the BG / NBD model and trained to predict the member login retention probability.
[0043] When APP conducts coupon distribution, the main way for members to receive coupons is through login-based coupon pop-up recommendations and active receipt in the coupon center. If members do not log in, even if coupons are issued, members will not receive the stimulus of coupons. Therefore, for member login loss prediction, different marketing strategies can be developed for members in different active states. BG / NBD model is a classic model for predicting customer lifetime value and customer churn, commonly used in customer relationship management and marketing fields. BetaGeometric is used to predict customer churn. The probability of each user losing after logging in each week follows the geometric distribution GeometricDistribution with parameter p (churn rate), and the PDF function is as follows:
[0044] ,
[0045] Where P is x, the number of weeks in the past year. If a member loses (p) in the xth week, he must survive (1-p) in the previous x-1 weeks. If p is small, the member's survival probability decreases slowly. The overall user churn rate p follows the BetaDistribution with shape parameters a and b, for example, the churn rate p of 100 users follows the beta distribution with a=1.0 and b=2.5.
[0046] In most scenarios, people apply BG / NBD model to use member transaction data to predict member order churn rate and CLV. In our login loss prediction, we need to replace transaction data with login records. Calculate the member characteristics recency / T / frequency in weeks.
[0047] We fit the member login characteristic data recency / frequency / T of BaiLian to learn the beta distribution parameters a / b through MLE.
[0048] Step S3, a member target category order probability prediction model is constructed based on the lightGBM classifier model and trained.
[0049] Where the features of the model include the following steps:
[0050] Step S301, all model features include member characteristics and combinations of characteristics in target format characteristics.
[0051] There are 80 member characteristics, which can be divided into several categories, including:
[0052] 1. Basic features 5, including gender, age, registration days, registration format, member level;
[0053] 2. Member cdp image 5, such as rfm, life cycle, etc.;
[0054] 3. Member online behavior 36, including search behavior, order behavior (number of ordered goods, category number, brand number, first-level category distribution, actual payment amount), browsing behavior (goods, coupon center), shopping cart behavior.
[0055] 4. Coupon-related features 26, including categories: number of coupons, number of coupons used; full reduction threshold, deduction amount, cancellation rate; coupon usage speed after receiving; coupon format, channel.
[0056] 5. Member offline behavior 8, such as parking frequency, license plate number, department store offline purchase frequency, supermarket offline purchase frequency.
[0057] Member's features for target format category 53, including:
[0058] 1. Member's online behavior features for target category and format 35, such as browsing behavior (goods, coupon center), order behavior (number of goods, category number, first-level category distribution, actual payment amount, recent consumption time), shopping cart behavior.
[0059] 2. Member's coupon features when consuming in target category 16, such as full reduction threshold, deduction amount, cancellation rate, coupon usage speed after receiving.
[0060] 3. Member's offline consumption behavior for target category 9, such as offline order number, order amount, recent consumption time, number of stores in United Supermarket.
[0061] Step S302, screening and determining the feature service of different prediction targets;
[0062] Model features include member features and member features for target category. Member single-dimensional features are the same, and member features for format category need to be aggregated to member dimension. By screening target category and format, member features for target category are determined.
[0063] Step S303, according to the requirements of lightGBM model, the features are cleaned, and the numerical value data is supplemented with 0; the category data is encoded based on OrdinalEncoder.
[0064] Step S304: Feature filtering is performed using sklearn's SelectKBest. If the target y is a binary classification task, and feature x is a categorical variable, a chi-square test is used; if x is a numerical variable, an ANOVA test is used. Significant variables with p-values < 0.05 are selected.
[0065] The above features are updated daily. Based on the data tracking details of the past year, the daily membership data is calculated, along with the membership-category features, and inserted into the daily partition. Alternatively, all membership features for a specific day can be calculated using a time parameter configuration.
[0066] Next, we set the target range for predicted orders. The predicted order range is a combination of different business formats (main site, home delivery, cloud store) / categories / amount ranges. Also, because the business involves coupon marketing, the coupon's applicable scope may be a combination of several business formats / categories. For example, we want to predict the probability of orders placed in the next 10 days from the main site or home delivery business formats, and from the fresh produce and food categories, within the range of 60-1000.
[0067] Because the number of positive and negative sample data points for members placing orders in the target category in the next N days differs significantly from the total number of users being calculated, potentially reaching 1 / 30 in extreme cases, the smote algorithm is used to upsample the positive samples to achieve a positive-to-negative sample ratio of 1 / 8.
[0068] The model training process is divided into offline testing and online training on the full dataset. Offline training and testing involves dividing the training data into 80% training and 20% testing. Performance metrics, including MSE, precision, and recall, are calculated on the testing data. The training prediction results are then visualized. Finally, the hyperparameters of lightGBM are obtained.
[0069] The full-data training (online) is based on the full-data train_xy training of the lightGBM classifier, and is currently being conducted daily.
[0070] Step S4: Predict the probability of a member placing an order, and perform bucketing to verify the recall rate and precision of the actual orders.
[0071] Using the hyperparameters of the model obtained earlier, we predict member ordering behavior to obtain the probability of a member placing an order. Members are then sorted from highest to lowest ordering probability. Assuming each bucket contains 5000 members, the predictions are grouped into buckets of 5000 based on probability. The prediction results are validated over the next N days. After predicting the ordering probability today, the buckets are compared with actual ordering members in the next 3 / 5 / 7 days to obtain the member click-through rate, recall rate, and precision rate.
[0072] Step S5: Develop a coupon distribution strategy based on order placement probability, customer CLV, and login retention probability.
[0073] Step S6, according to the formulated coupon distribution strategy, the marketing activities are carried out, and the member login retention estimation model and the member target category order probability prediction model are optimized and adjusted according to the activity results.
[0074] The important features for different categories are analyzed by tools to correct the model, and the tools include FeatureImportance function of lightGBM itself or SHAP value.
[0075] The FeatureImportance function of lightGBM itself can obtain the importance of the feature, which is from the total gain brought by the decision tree used for segmentation. For example, for the food and fresh product with high repurchase, the near-term consumption category features and the near-term purchase behavior of the member are mainly relied on. For durable goods, the model pays more attention to the recent search and purchase behavior of the target category. For products with long repurchase cycle, more attention is paid to the long-term consumption behavior of the customer.
[0076] The model in the technical solution is used to predict the order probability. In general, the mean of the probability distribution of the predicted purchase is higher than that of the non-purchase. The model can distinguish between people who actually purchase and those who do not. However, due to the nature of the categories, the prediction accuracy and certainty of different categories are different. According to the offline test results of the model, the food and fresh category has the highest prediction accuracy, followed by the shoe and clothing purchase prediction of the main station + cloud store. The purchase prediction of the main station 3c home appliance category is relatively difficult, but still some people can be accurately identified.
[0077] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. An incentive method based on a user category purchase prediction model, characterized in that, Includes the following steps: Step S1: Determine the marketing objectives for the business coupons, including the business format, marketing category, membership scope, amount range, and activity duration. Step S2: Construct a member login retention prediction model based on the BG / NBD model, train it, and predict the member login retention probability. Step S3: Construct and train a member target category order probability prediction model based on the lightGBM classifier model; Step S4: Predict the probability of a member placing an order, and perform bucketing to verify the recall rate and precision of the actual order placement. Step S5: Develop a coupon distribution strategy based on order placement probability, customer CLV, and login retention probability. Step S6: Conduct marketing activities according to the established coupon distribution strategy, and optimize and adjust the member login retention prediction model and the member target category order probability prediction model based on the activity results.
2. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, In step S2, the member churn rate p follows a Beta distribution with shape parameters a and b. By fitting the feature data recency / frequency / T of Bailian member logins using MLE, the parameters a and b of the Beta distribution are learned.
3. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, In step S3, the positive samples are upsampled using the SMOTE algorithm to keep the ratio of positive to negative samples at 1:
8.
4. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, The method for determining the features of the model in step S3 is as follows: Step S301: The model features include the member's own features and the combination of features in the target business category. Step S302: Filter and determine feature services for different prediction targets; Step S303: Clean the features according to the requirements of the lightGBM model; Step S304: Filter features based on sklearn's SelectKBest.
5. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, The member target category order probability prediction model calculates the characteristics of members and their categories on a daily basis, based on the data tracking details from the past year. Update the features and insert the current day's partition.
6. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, The member target category order probability prediction model predicts all member characteristics on a future day by configuring time parameters.
7. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, In step S3, the model is tested offline and trained online on the full dataset, respectively. Offline testing involves dividing the training data into two parts, with 80% used for training and 20% for testing. The optimal hyperparameters of the model are obtained through training. The full-data online training of the model involves training the lightGBM classifier daily using the full dataset.
8. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, In step S4, the bucket verification first uses the model parameters selected from offline testing to predict the probability of a member placing an order for a product category in the next N days. The members are then sorted and bucketed according to their order probability, and the order probability of each bucket is verified.
9. The incentive method based on a user category purchase prediction model according to claim 1, characterized in that, The model was refined by analyzing key features for different product categories using tools, including the FeatureImportance function built into lightGBM and SHAP value.
10. An incentive method based on a user category purchase prediction model, characterized in that, This includes modules for marketing objectives, retention prediction, order prediction, bucket verification, campaign delivery, and optimization. The marketing objectives module is used to determine the marketing objectives for business coupons. The marketing objectives include business type, marketing category, membership scope, amount range, and activity duration. The retention prediction module is used to build a member login retention prediction model based on the BG / NBD model, train it, and predict the member login retention probability. The order prediction module is used to build and train a member target category order probability prediction model based on the lightGBM classifier model; The bucket verification module is used to predict the probability of a member placing an order and to perform bucket verification on the recall and precision of the actual orders. The campaign delivery module is used to formulate coupon delivery strategies based on order placement probability, customer CLV, and login retention probability. The optimization module is used to conduct marketing activities according to the established coupon distribution strategy, and to optimize and adjust the member login retention prediction model and the member target category order probability prediction model based on the activity results.