Research on Commodities Constraint Optimization Based on Graph Neural Network Prediction

Abstract

Business intelligence makes good sale prediction crucial in any commercial activity as it has a significant impact on production and supply plan. However, practical commercial data presents explicit constraints, that how to get the optimal forecasts of commodity sales under the constraints is a vital problem many researchers face. The present research proposes a prediction model which combines graph convolution neural network and node bipartite graph. Firstly, the node bipartite graph algorithm is used to merge the constraint graph and the store graph, obtaining the “store-constraint bipartite graph”. Secondly, a graph convolutional neural network integrating GRU and AR is utilized to extract temporal features (X). Finally, a fully connected network is applied to predict the optimal solution (Y) after constraint optimization. The former can effectively learn complex features of stores, meanwhile, the later combines the constraint conditions with the store, which can effectively predict the sales of goods under the constraint conditions. In terms of model performance, we compared the proposed model with the classical method such as SVR, LSTM, ARIMA. RMSE, MSE, MAE and MAPE are used for evaluation indexes, and the results show that MAPE for one month’s sales of some product from both datasets is 7.75%.

Keywords

• Graph convolutional networks
• constraint optimization
• graph representation learning
• sales prediction
• business intelligence