Proactive Model Adaptation Against Concept Drift for Online Time Series Forecasting

TL;DR

This work tackles concept drift in online time series forecasting under ground-truth delays by introducing Proceed, a proactive adaptation framework. Proceed estimates a high-dimensional drift representation between recent training data and the current test input, then maps this drift to compact parameter shifts via a bottleneck-based adaptation generator trained on synthetic concept drifts. Empirical results across five real-world datasets show that Proceed consistently outperforms online baselines and substantially improves forecasting accuracy while maintaining efficiency. The approach offers a practical pathway to resilience against drift in settings with delayed supervision, and code is released for reproducibility.

Abstract

Time series forecasting always faces the challenge of concept drift, where data distributions evolve over time, leading to a decline in forecast model performance. Existing solutions are based on online learning, which continually organize recent time series observations as new training samples and update model parameters according to the forecasting feedback on recent data. However, they overlook a critical issue: obtaining ground-truth future values of each sample should be delayed until after the forecast horizon. This delay creates a temporal gap between the training samples and the test sample. Our empirical analysis reveals that the gap can introduce concept drift, causing forecast models to adapt to outdated concepts. In this paper, we present Proceed, a novel proactive model adaptation framework for online time series forecasting. Proceed first estimates the concept drift between the recently used training samples and the current test sample. It then employs an adaptation generator to efficiently translate the estimated drift into parameter adjustments, proactively adapting the model to the test sample. To enhance the generalization capability of the framework, Proceed is trained on synthetic diverse concept drifts. Extensive experiments on five real-world datasets across various forecast models demonstrate that Proceed brings more performance improvements than the state-of-the-art online learning methods, significantly facilitating forecast models' resilience against concept drifts. Code is available at https://github.com/SJTU-DMTai/OnlineTSF.

Figures (7)

• Figure 1: (a) Example of feedback delay when $H=3$. At online time $t=5$, we use observation values $\mathbf X_5=\{\mathbf v_1,\cdots,\mathbf v_5\}$ to forecast future values $\mathbf Y_5=\{\mathbf v_6,\mathbf v_7,\mathbf v_8\}$, while the ground truth is known until $t=8$ and the feedback of forecasting performance arrives with 3-step delay. (b) A temporal gap always exists between new training samples $\mathcal{D}_{t-}$ and the test sample $\mathbf X_t$, where concept drift may occur.
• Figure 2: Illustration of the mapping from time series data change to parameter shift. Concept encoders $\mathcal{E}$ and $\mathcal{E}'$ encode time series into concept representations. Adaptation generator $\mathcal{G}$ decodes the conceptual difference into parameter shift.
• Figure 3: Example of synthetic concept drifts in historical training data and online concept drifts when $H=2$. The circles represent concepts, while the arrows represent concept drifts.
• Figure 4: Overview of the model adapter in Proceed that comprises two concept encoders and an adaptation generator.
• Figure 5: Efficiency comparison on the Traffic dataset ($H=24$). The horizontal axis is the online forecasting latency (millisecond) between updating the model and obtaining online predictions. The vertical axis is the average MSE on test data. The size of each circle represents the peak amount of GPU memory occupation (GB).

Theorems & Definitions (2)

• Definition 1: Time Series Forecasting
• Definition 2: Online Model Adaptation