Tackling Time-Series Forecasting Generalization via Mitigating Concept Drift

TL;DR

This paper addresses the challenge of distribution shifts in time-series forecasting by distinguishing temporal shift and concept drift, with the latter being underexplored. It introduces soft attention masking (SAM) to extract invariant patterns from both lookback and horizon exogenous features, forming a surrogate feature $X^{SUR}$ and modeling $P(Y^H|X^{SUR})$ to improve generalization. Recognizing temporal-shift mitigation as a prerequisite, the authors propose ShifTS, a model-agnostic framework that normalizes marginals, forecasts $X^{SUR}$, and then predicts $Y^H$ via an aggregation step, optimizing $\mathcal{L}=\mathcal{L}_{SUR}+\mathcal{L}_{TS}$. Comprehensive experiments on six real-world datasets across multiple models show that ShifTS consistently improves forecasting accuracy and often outperforms both concept-drift and temporal-shift baselines. The approach offers a practical, generalizable strategy for robust time-series forecasting under evolving data distributions.

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.

Figures (6)

• Figure 1: Comparison between conventional time-series forecasting and our approach. Our approach identifies invariant patterns in lookback and horizon window as $\bm{\mathrm{X}}^{SUR}$ and then models a stable conditional distribution accordingly to mitigate concept drift.
• Figure 2: Diagram of ShifTS, consisting of three components: (a) normalization at the start (c) denormalization at the end to address temporal shifts, and (b) a two-stage forecasting process-The first stage predicts surrogate exogenous features, $\bm{\hat{\mathrm{X}}}\newline^{\mathrm{SUR}}$, identified by the SAM, which capture invariant patterns essential for forecasting the target; The second stage uses both the predicted surrogate exogenous features and the original $\mathrm{Y}^L$ to predict $\mathrm{Y}^H$.
• Figure 3: Left (a): The performance gains of ShifTS versus the mutual information shared between $\bm{\mathrm{X}}^H$ and $\bm{\mathrm{Y}}^H$. Greater mutual information in $\bm{\mathrm{X}}^H$ compared to $\bm{\mathrm{Y}}^H$ correlates with more significant performance gains achieved by ShifTS. Right (b): Ablation Study. Addressing either concept drift or temporal shift individually provides certain benefits in forecasting accuracy. ShifTS that tackles both achieves the lowest forecasting error.
• Figure 4: Demonstration of temporal shift phenomenon within time series data, showcasing the variations in statistical properties, including mean and variance, over time as the emergence of temporal shift (Red: ground truth; Yellow: N-BEATS prediction; Blue: N-BEATS+RevIN prediction).
• Figure 5: Demonstration of concept drift phenomenon within time series data, showcasing the variations in correlation structures between target series $\bm{\mathrm{Y}}$ and exogenous feature $\bm{\mathrm{X}}$ over time as the emergence of concept drift (Red: ground truth; Yellow: N-BEATS prediction; Blue: N-BEATS+RevIN prediction).

Theorems & Definitions (2)

• Definition 2.1: Temporal Shift shimodaira2000improvingdu2021adarnn
• Definition 2.2: Concept Drift lu2018learning