$spadesuit$ SPADE $spadesuit$ Split Peak Attention DEcomposition

2024-11-06

AI Summary

To address demand forecasting overshoot during peak events (e.g., promotions, holidays) and subsequent post-peak prediction bias, this paper proposes a decoupled forecasting framework that decomposes demand into two subtasks: peak-event and non-peak-event modeling. We introduce a novel hybrid architecture integrating masked convolution and peak attention: the masked convolution filters out peak-induced noise by suppressing interference outside temporally bounded event windows, while the peak attention module dynamically captures the evolving temporal influence of events. A dual-path temporal decomposition further enhances feature disentanglement. Evaluated on a global retail dataset comprising billions of SKUs, our method reduces post-peak forecasting error by 4.5% and improves peak-period accuracy by 3.9%, significantly outperforming state-of-the-art end-to-end time-series models.

Technology Category

Planning, Routing, and Scheduling: Optimization of Spatio-temporal SystemsMachine Learning: Time-Series/Data StreamsSearch and Optimization: Non-convex Optimization

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systemsGraph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphs

Abstract

Demand forecasting faces challenges induced by Peak Events (PEs) corresponding to special periods such as promotions and holidays. Peak events create significant spikes in demand followed by demand ramp down periods. Neural networks like MQCNN and MQT overreact to demand peaks by carrying over the elevated PE demand into subsequent Post-Peak-Event (PPE) periods, resulting in significantly over-biased forecasts. To tackle this challenge, we introduce a neural forecasting model called Split Peak Attention DEcomposition, SPADE. This model reduces the impact of PEs on subsequent forecasts by modeling forecasting as consisting of two separate tasks: one for PEs; and the other for the rest. Its architecture then uses masked convolution filters and a specialized Peak Attention module. We show SPADE's performance on a worldwide retail dataset with hundreds of millions of products. Our results reveal a reduction in PPE degradation by 4.5% and an improvement in PE accuracy by 3.9%, relative to current production models.

Problem

• Demand Forecasting
• Seasonal Variability
• Prediction Accuracy

Innovation

• SPADE model
• Split Peak Attention
• demand peak handling