Real-Time Data Pipelines for Demand Forecasting in Retail Paint Distribution Networks

Abstract

Retailers are provided with a large quantity of historical demand time-series data, and they need to make demand predictions in order to maintain sufficient inventory and to prepare supply and logistics, etc. In particular, large-scale retail paint companies adopt decentralized distribution networks, which can be viewed as an unbalanced directed graph. A central warehouse supplies paints to multiple regional warehouses and shops. Different from traditional time-series forecasting problems, in addition to the time-series data, other relational features on the directed graph are also provided that potentially are very powerful. However, demand prediction with complex relational features and graphs over a large scale is very challenging. Probabilistic forecasting of time-series with complex relational structure is rarely studied.

This work presents an innovative framework called GRIT, Graph attention-based GNN based probabilistic Forecasting Pipeline for Demand in Retail Paint distribution network. Browsing the architecture of GRIT, a Gated Relational Graph Convolutional module predicts the demand with the graph neural network based on attention mechanism. The proposed GNN model is very suitable for modeling relational and heterogeneous features and solves how the distributions of related time-independent variables on the directed graph are beneficial for demand prediction. An asymmetric penalized gradient boosting tree model produces point predictions and achieves the best computational efficiency. An ingenious parallel pipeline is designed to accelerate the inference speed of the final model for high-frequency predictions. Based on latency simulation experiments, how the multi-query character of the forecasting task is beneficial to processing large quantities of time-series is analyzed with detailed experiments for demand queries at highly skewed lags against time. Finally, a completely synthetic production scenario is presented to evaluate the performance of forecasting on serial product demands, which is very useful for the advertising and stocking of paint distribution.

Demand of paint prediction in multi-hierarchy paint distribution networks accounts for a novel scenario in time-series forecasting with complex relations globally and locally. GNN techniques are newly proposed for the probabilistic forecasting task with complex relational features, which is rarely studied. A generic pipeline is provided to enhance forecasting speed and scale with useful insights in latency analysis. In addition, a synthetic experimental scenario is proposed to assess the forecasting performance of multi-aggregation serial time-series, which inspires multi-aggregation modeling frameworks