多特征加权集成模型在用户购买预测中的研究  

Research on Multi-Feature Weighted Ensemble Model in User Purchase Prediction

作　　者：胡静静 樊军[1] HU Jing-jing;FAN Jun(School of Mechanical Engineering,Xinjiang University,Xinjiang Urumqi 830047,China)

机构地区：[1]新疆大学机械工程学院,新疆乌鲁木齐830047

出　　处：《机械设计与制造》2024年第4期373-378,383,共7页Machinery Design & Manufacture

基　　金：国家自然科学基金地区科学基金项目(11462021)。

摘　　要：为提升用户线上购买行为预测效果,对用户-商品线上交互等数据通过函数拟合、计数、加和、均值、比值、设置非数值类别型特征、二次组合衍生等方法构造提取3个特征群,建立了针对此种业务场景的特征工程。提出一种基于堆叠(Stacking)的加权异质集成模型,将Stacking集成框架中第一层异质基分类器在数据集上的性能排序信息转化为一组约束,添加到第二层LPBoost算法中,求解改进的LPBoost算法目标规划问题得到基分类器更佳组合权重,构建加权集成模型预测用户购买行为。选用阿里云天池官方发布的用户行为数据集进行实验验证,得到8.51%的F1值,优于对比方案。In order to improve the prediction effect of users’online purchase behaviors,3 groups of features are constructed and ex-tracted from the user-products online interaction data by fitting with function,counting,adding,calculating the mean,calcu-lating the ratio,setting non-numeric categorical features,combining and deriving once,etc.The feature engineering for this business scenario is set up,and a weighted heterogeneous ensemble model based on Stacking is proposed.The performance rank-ing information of the first-layer heterogeneous base classifiers of the Stacking ensemble framework on the dataset is transformed into a set of constraints,which are added to the LPBoost algorithm in the second-layer of the stacking ensemble framework.The goal programming problem of the improved LPBoost algorithm is solved to obtain the better combination weight of the base classi-fiers,and the weighted ensemble model is constructed to predict the user’spurchase behavior.The user behavior data set officially released by Alibaba Cloud Tianchi is used for experimental verification,and the F1 value of 8.51%is obtained,which is superior to comparison schemes.

关 键 词：购买行为预测 特征构造 特征工程 Stacking加权集成 

分 类 号：TH16[机械工程—机械制造及自动化] T-9[一般工业技术]