Explainability Challenges in Continual Learning for Time Series Forecasting

Deep learning models are powerful for time series forecasting, but their application in real-world environmental monitoring is hampered by non-stationary data dynamics and a lack of explainability. This study delves into the role of explainability in understanding continual learning within adaptive time series forecasting, specifically focusing on strategies like Experience Replay. The researchers examined neural forecasting architectures such as PatchMixer, PatchTST, and DLinear, enhanced with attention-based sampling mechanisms to facilitate model adaptation over time. They employed explainability techniques like attention rollout and Grad-CAM to analyze both the predictive behavior of the models and their sampling strategies within a continual learning framework. Experiments conducted on real-world piezometric time series, which exhibit heterogeneous patterns and regime shifts, provided valuable insights. The analysis of model and sampling behaviors highlighted the challenges and opportunities of using explainability to comprehend continual learning dynamics, showing how attribution patterns evolve and inform data selection in changing forecasting scenarios.