Hierarchical Spatio-Temporal Neural Network For Sales Prediction

As a traditional time series problem,sales prediction is widely used in industries such as manufacturing,retail,marketing,wholesaling and supply chain.Sales prediction is essential for managers and helps in inventory management,supply and demand analysis,business strategy and economic decision making.With the development of machine learning,more studies focus on exploring the rich variety of structural factors that affect sales prediction,such as product brand,ingredients,etc.,or external characteristics related to the product itself,such as date,promotion,holidays,etc.Most works usually treat the sales prediction problem as a multivariate time series forecasting problem,and very few of them take product-product effects into account.However,researches showed that the complex correlation of products with external factors has an ambiguous but indispensable impact on the prediction of sales.In the absence of transaction records and user behavior,related works calculate the relevance of commodities based on the sales records of products.The method of calculating item sales records one-by-one results in an increase in the amount of calculation as the number of items increases,and re-calculation of associations is required as the sales records grow.Therefore,it is urgent to define product associations and obtain this association.In addition,related works use the established rules for filtering products to reduce the number of products used in the research,thereby reducing the calculation amount of sales prediction correlation analysis.However,this method ultimately utilizes less related products and cannot be applied to large-scale product data.Therefore,how to ensure the large-scale utilization of related products remains as a problem.In order to handle these two problems,this paper proposes a hierarchical spatio-temporal graph neural networks for sales prediction,which builds product related information with the help of external data,and make full use of related product data to improve the performance of sales forecasting.Specifically,this paper derives product relationships from external knowledge,so that the products can be predicted on a large scale and more relevant products can be used to assist the prediction.The hierarchical spatio-temporal graph neural network constructed in this paper can make full use of the time series information of related products,and simultaneously make fine use of the hierarchical categories of products.In addition,this paper constructs multi-modal data of products from external knowledge and datasets to help modeling products,and utilizes attention mechanism to fuse multi-modal data and time series data.On the whole,this paper uses hierarchical spatio-temporal graphs to fuse time series data of related products,and simultaneously fuses multi-modal data and time series representations of spatio-temporal graphs to improve the performance of sales forecasting.In this paper,experiments are conducted on a real-world sales dataset.The results show that the proposed sales forecasting model based on a hierarchical spatio-temporal graph neural network exceeds the benchmark methods that do not utilize associated product information,which proves the effectiveness of the proposed method.Additionally,this paper set up ablation experiments to verify the effectiveness of associated products information,multimodal data fusion,and hierarchical spatio-temporal graph modules,respectively.