Peak period demand forecasting with proxy data: GNN-enhanced meta-learning

Demand prediction is a crucial task for e-commerce and physical retail businesses, especially during high-stake sales events. However, the limited availability of historical data from these peak periods poses a significant challenge for traditional forecasting methods. In this paper, we propose a novel approach that leverages proxy data from non-peak periods, enriched by features learned from a graph neural networks (GNNs) based forecasting model, to predict demand during peak events. We formulate the demand prediction as a meta-learning problem and introduce the Feature-based First-Order Model-Agnostic Meta-Learning (F-FOMAML) algorithm, which adapts to new tasks by conditioning on the GNN-generated relational metadata. Empirical evaluations on large-scale industrial datasets demonstrate the superiority of our approach, with our model consistently outperforming state-of-the-art base-lines in the demand prediction task, by 26.24% on the in-ternal vending machine dataset and 8.7% on the public JD.com dataset over the Mean Absolute Error.