Stacking集成学习方法在销售预测中的应用  

APPLICATION OF STACKING INTEGRATED LEARNING METHOD IN SALES FORECASTING

作　　者：王辉[1] 李昌刚[1] Wang Hui;Li Changgang(School of Information and Intelligence,Zhejiang Wanli University,Ningbo 315000,Zhejiang,China)

机构地区：[1]浙江万里学院信息与智能学院,浙江宁波315000

出　　处：《计算机应用与软件》2020年第8期85-90,共6页Computer Applications and Software

基　　金：宁波市科技特派员团队项目(2018C80002-10)。

摘　　要：为了提高单一预测模型在销售预测中的性能,提出一种在多机器学习模型融合下基于Stacking集成策略的销售预测方法。将数据划分为四个同分布的数据集;基于各数据集训练多个基学习器;以XGBoost算法为元学习器构建两层Stacking集成学习方法;使用德国Roseman超市在Kaggle平台上的销售数据对算法进行验证。实验结果表明:在Stacking模型中,元学习器利用各基学习器的算法优势提升了模型的预测性能,相比单个模型在测试集上的均方根百分误差,Stacking模型最高减少了23.5%,最低减少了1.8%。In order to improve the performance of single prediction model in sales forecasting,we propose a sales forecasting method based on Stacking integration strategy under the fusion of multi-machine learning model.The data was divided into four equally data sets with the same distribution;multiple base learners were trained based on each data set;the two-layer Stacking integrated learning method was constructed with XGBoost algorithm as the meta-learner;we verified the algorithm using the sales data from the German Roseman supermarket on the Kaggle platform.The experimental results show that in the Stacking model,the meta-learner improves the prediction performance of the model by taking advantage of the algorithm advantages of each base learner.The Stacking model reduces the error by 23.5%at the highest and 1.8%at the lowest,compared with the root mean square error of a single model on the test set.

关 键 词：机器学习 销售预测 Stacking集成学习 XGBoost 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]