Research on Forecasting Sales of Products Based on Spatiotemporal Graph Neural Network

Abstract

Intelligent marketing and recommendation are a core business of commercial companies, and accurate prediction of sales is the premise and foundation for greater efficiency of smart marketing and recommendation. In order to predict product sales, deep neural network (DNN), convolutional neural network (CNN), time series analysis and other methods have been put forward, but most of which only focus on the temporal or spatial characteristics of data. According to modeling and analyzing sales of products, they are closely related to the spatial location and time of the corresponding merchants. The goal is to predict the sales of products accurately at a given time and place, we advance a hybrid model of CNN-LSTM to forecast sales. Firstly, a large-scale knowledge graph system based on merchants is constructed, which describes the sales data and the relevant interaction scenarios of the corresponding business, merchants and users through the data model of a graph, and add the spatial and data characteristics of the business data on the graph model to describe the temporal and spatial characteristics of the merchants. Based on the constructed business knowledge graph, graph convolutional neural network (GCN) is used to aggregate information and obtain spatial features. Correspondingly, long short-term memory (LSTM) is used to extract time features. Researchers combine the two characteristics to make the sales forecast. In this study, neural network and GCN-LSTM algorithm are respectively used to carry out experiments on two kinds of product regulations. The result shows that the sales predicted by hybrid model of GCN-LSTM is almost as equal as the actual sales. The average accuracy of the proposed model is 89%.