When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate

Abstract Time series extrinsic regression (TSER) aims to predict a continuous target variable from an input time series. It arises in domains such as healthcare, finance, environmental monitoring, and engineering, where both accuracy and interpretability are essential. Despite strong predictive performance, state-of-the-art TSER models typically operate as black boxes, making it difficult to understand which temporal patterns drive their predictions. Post-hoc explanation methods attempt to address this issue but often produce coarse, noisy, or unstable explanations. Inherently interpretable approaches based on additive decompositions or concepts offer an alternative, yet they often require concept supervision, struggle to capture multivariate interactions, and lack expressiveness for complex temporal patterns. To address these limitations, we propose Mask-and-Aggregate Networks for Time Series (MAGNETS), an inherently interpretable neural architecture for TSER. MAGNETS learns a compact set of human-understandable concepts directly from data, without requiring concept annotations. Each concept corresponds to a learned, mask-based aggregation over selected input features, revealing both which features influence the prediction and when they matter. Predictions are expressed as transparent combinations of these concepts, yielding faithful, input-specific explanations by design. Experiments on synthetic and real-world univariate and multivariate TSER datasets show that MAGNETS achieves accuracy comparable to black-box models while substantially outperforming existing interpretable baselines, particularly on tasks involving multivariate feature interactions.