基于数据挖掘的超市商品销量预测

Supermarket Commodity Sales Forecast Based on Data Mining

摘要: 针对超市商品短时间内销量预测问题，本文通过对比几种基本模型，提出了一种基于LightGBM和支持向量回归模型相结合的预测模型。该模型不仅通过对用户的行为数据进行量化特征提取和商品属性的特征提取，同时结合了时间滑动窗口在特征处理上的优势，将商品的销售数据作为前后关联数据进行动态特征提取，再通过多模型关系的融合，对商品数据进行预测。实验结果显示，经过滑窗法特征提取后，通过对比支持向量回归模型和LightGBM预测模型，发现LightGBM预测模型效果略优于支持向量回归模型，通过组合支持向量回归模型和LightGBM模型，发现超市销量预测模型的均方根误差仅为1.23209，明显高于单模型预测结果。因此，该模型是短期超市商品销量预测的一种有效方法。

Abstract: Based on the comparison of several basic models, a prediction model based on LightGBM and support vector regression model is proposed in this paper. This model not only extracts the features of the user’s behavior data and the features of commodity attributes, but also combined with the advantages of time sliding window in feature processing, extracts dynamic features by using the sale data of the commodity and correlation data, and then uses the fusion of multiple models to predict the commodity data. The experimental results show that after the feature extraction of sliding window method, by comparing support vector regression model and LightGBM prediction model, it is found that the effect of LightGBM prediction model is slightly better than the support vector regression model. By combining the support vector regression model and the LightGBM model, the root-mean-square error of the supermarket sales forecast model is 1.23209, which is significantly higher than the single model prediction results. Therefore, this model is an effective method to predict the sales volume of short-term supermarket.