融合集成算法与宽度学习的商品需求量预测

Fusing Integrated Algorithms with Broad Learning System for Commodity Demand Forecasting

由于电商订单销售数据中全品类商品的种类繁多、分类多层化,且数据集还存在时序长度分布不均、地区差异、定性变量及非线性特征变量的处理等问题,导致需求量的预测任务较困难。为了解决上述问题,本研究提出一种宽度学习集成框架,将机器学习中的Random Forest、GBDT、XGBoost和LightGBM与宽度学习模型进行随机融合,并分别进行验证,对比模型效果。经实证分析结果表明:LightGBM-BLS模型具有最优的预测性能和计算性能,它在保持LightGBM模型计算优势的同时,大幅度地提升了模型本身的预测精度,使拟合优度达到0.99,评价指标RMSE、MSE降低90%以上,MAE降低85%以上。

Due to the wide variety of full-category commodities in e-commerce order sales data, multi-layered categorization, and the dataset also has the problems of uneven distribution of time-series lengths, regional differences, and the treatment of qualitative variables and non-linear feature variables, which leads to a more difficult task of demand prediction. To solve the above problems, this study proposes a breadth learning integration framework, which stochastically fuses Random Forest, GBDT, XGBoost and LightGBM in machine learning with the breadth learning model, and validates and compares the model effects respectively. Empirical analysis results show that the LightGBM-BLS model has optimal prediction performance and computational performance, which maintains the computational advantages of the LightGBM model while substantially improving the prediction accuracy of the model itself, so that the goodness of fit reaches 0.99, and the evaluation indexes of RMSE and MSE are reduced by more than 90%, and MAE is reduced by more than 85%.