基于Stacking集成学习的服装网络直播销量预测

Online live broadcast sales forecast for clothing based on Stacking integrated learning

为了解决单一模型对直播销量预测效果不佳的情况,提出利用Stacking集成学习模型对4种单一机器学习模型进行融合。利用Spearman相关性分析和3种树模型的特征贡献度来进行特征选择,选用网格搜索以及贝叶斯优化算法进行模型参数选择。利用抖音直播李维斯品牌牛仔裤品类数据对算法进行实例验证。对比不同组合模型的MAE、MSE、RMSE和SMAPE值,实验证明:选用随机森林、支持向量回归、Xgboost为基学习器,线性回归为元学习器的两层Stacking集成学习模型对服装网络直播销量的预测效果优于单一机器学习模型以及其他组合模型,SMAPE的误差较单一模型最高下降6.97%,最低下降2.53%。

In order to solve the problem that a single model cannot forecast live broadcast sales,a stacking integrated learning model that fuses four separate machine learning models is presented.The Spearman correlation analysis and feature contribution of three tree models are used for feature selection,and the grid search and Bayesian optimization algorithm are used to select model parameters.The algorithm is verified by the category data of Levi’s jeans on Douyin.By comparing the MAE,MSE,RMSE and SMAPE values of a single model,the experiment proves that the two-layer Stacking ensemble learning model using random forest,support vector regression,and Xgboost as the base learner and linear regression as the meta-learner is better than a single machine learning model and other combined models in the online live broadcasts sales forecasting effect of clothing.Compared with the single model,the error of SMAPE decreases by 6.97% at the highest level and 2.53% at the lowest level.