Tackling Time-Series Forecasting Generalization via Mitigating Concept Drift

2025-10-16

AI Summary

Time series forecasting suffers from deteriorating generalization due to temporal distributional shifts, with existing methods primarily addressing temporal shift while inadequately modeling concept drift. This paper proposes ShifTS, the first unified framework that explicitly distinguishes and jointly models both shifts in time series forecasting. ShifTS adopts a two-stage, model-agnostic approach: first mitigating temporal shift via soft attention and temporal alignment; then correcting concept drift through invariance learning. Crucially, it requires no architectural modifications to underlying forecasting models. Extensive experiments across multiple real-world datasets demonstrate that ShifTS consistently enhances both robustness and accuracy of state-of-the-art forecasting models. It outperforms all baseline methods—including those handling only one type of shift or naively combining both—across diverse evaluation metrics and settings.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Planning, Routing, and Scheduling: Temporal Planning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Time-series forecasting finds broad applications in real-world scenarios. Due to the dynamic nature of time series data, it is important for time-series forecasting models to handle potential distribution shifts over time. In this paper, we initially identify two types of distribution shifts in time series: concept drift and temporal shift. We acknowledge that while existing studies primarily focus on addressing temporal shift issues in time series forecasting, designing proper concept drift methods for time series forecasting has received comparatively less attention. Motivated by the need to address potential concept drift, while conventional concept drift methods via invariant learning face certain challenges in time-series forecasting, we propose a soft attention mechanism that finds invariant patterns from both lookback and horizon time series. Additionally, we emphasize the critical importance of mitigating temporal shifts as a preliminary to addressing concept drift. In this context, we introduce ShifTS, a method-agnostic framework designed to tackle temporal shift first and then concept drift within a unified approach. Extensive experiments demonstrate the efficacy of ShifTS in consistently enhancing the forecasting accuracy of agnostic models across multiple datasets, and outperforming existing concept drift, temporal shift, and combined baselines.

Problem

• Addressing concept drift in time-series forecasting generalization
• Identifying distribution shifts including temporal and concept drift
• Proposing unified framework to mitigate temporal and concept drift

Innovation

• Soft attention mechanism finds invariant time series patterns
• ShifTS framework tackles temporal shift before concept drift
• Method-agnostic approach enhances forecasting accuracy across datasets