Spatio-temporal multi-graph networks for demand forecasting in online marketplaces
Demand forecasting is fundamental to successful inventory planning and optimisation of logistics costs for online marketplaces such as Amazon. Millions of products and thousands of sellers are competing against each other in an online marketplace. In this paper, we propose a framework to forecast demand for a product from a particular seller (referred as offer/seller-product demand in the paper). Inventory planning and placements based on these forecasts help sellers in lowering fulfillment costs, improving instock availability and increasing shorter delivery promises to the customers. Most of the recent forecasting approaches in the literature are one-dimensional, i.e., during prediction, the future forecast mainly depends on the offer, i.e., its historical sales and features. These approaches don't consider the effect of other offers and hence, fail to capture the correlations across different sellers and products seen in situations like, (i) competition between sellers offering similar products, (ii) effect of a seller going out of stock for the product on competing seller, (iii) launch of new competing products/offers and (iv) cold start offers or offers with very limited historical sales data. In this paper, we propose a general demand forecasting framework for multivariate correlated time series. The proposed technique models the homogeneous and heterogeneous correlations between sellers and products across different time series using graph neural networks (GNN) and uses state-of-the-art forecasting models based upon LSTMs and TCNs for modelling individual time series. We have experimented with various GNN architectures such as GCNs, GraphSAGE and GATs for modelling the correlations. We applied the framework to forecast the future demand of products, sold on Amazon, for each seller and we show that it performs ~16% better than state-of-the-art forecasting approaches.