TS-Unlearn: A Dual-Objective Unlearning Framework for Time Series Forecasting

The growing importance of data privacy and regulatory compliance has made machine unlearning a critical requirement for modern predictive systems. However, existing unlearning methods are largely tailored to classification tasks and cannot be directly applied to time series forecasting, where models must handle continuous outputs, long-range temporal dependencies, and multivariate correlations. In this work, we propose TS-Unlearn, a principled and efficient unlearning framework specifically designed for time series forecasting models. We begin by formalizing unlearning in forecasting settings and derive a theoretically grounded forgetting objective based on an upper bound of the KL divergence, showing that forgotten predictions should converge toward a Gaussian noise distribution. To make this feasible for regression models, we justify replacing KL with an \(L_2\) surrogate via a mutual-information argument, enabling stable optimization. We further model unlearning as a dual-objective problem—forgetting and remembering—and introduce a Pareto-based dynamic weighting mechanism that adaptively balances the two during training without manual tuning. Extensive experiments on multiple real-world datasets demonstrate that our approach achieves strong and reliable unlearning, outperforming existing baselines while maintaining high predictive accuracy on retained and unseen data.