When, How Long and How Much? Interpretable Neural Networks for Time Series Regression by Learning to Mask and Aggregate

1 When, How Long and How Much? Interpretable Neural Networks for Time Series Regression by Learning to Mask and Aggregate Florent Forest, Amaury Wei, Olga Fink Abstract—Time series extrinsic regression (TSER) refers to the task of predicting a continuous target variable from an input time series. It appears in many domains, including healthcare, finance, environmental monitoring, and engineering. In these settings, accurate predictions and trustworthy reasoning are both essential. Although state-of-the-art TSER models achieve strong predictive performance, they typically operate as black boxes, making it difficult to understand which temporal patterns drive their decisions. Post-hoc interpretability techniques, such as feature attribution, aim to to explain how the model arrives at its predictions, but often produce coarse, noisy, or unstable expla- nations. Recently, inherently interpretable approaches based on concepts, additive decompositions, or symbolic regression, have emerged as promising alternatives. However, these approaches remain limited: they require explicit supervision on the concepts themselves, often cannot capture interactions between time-series features, lack expressiveness for complex temporal patterns, and struggle to scale to high-dimensional multivariate data. To address these limitations, we propose MAGNETS (Mask- and-AGgregate NEtwork for Time Series), an inherently inter- pretable neural architecture for TSER. MAGNETS learns a compact set of human-understandable concepts without requir- ing any annotations. Each concept corresponds to a learned, mask-based aggregation over selected input features, explicitly revealing both which features drive predictions and when they matter in the sequence. Predictions are formed as combinations of these learned concepts through a transparent, additive structure, enabling clear insight into the model’s decision process. Experiments on synthetic and real-world univariate and mul- tivariate TSER datasets show that MAGNETS closely matches the accuracy of black-box models while substantially outperform- ing existing interpretable baselines, particularly on multivariate datasets involving feature interactions. Finally, we also show that MAGNETS provides more faithful and informative explanations than post-hoc methods. The code implementation and datasets are publicly available at https://github.com/FlorentF9/MAGNETS. Index Terms—Time series regression, Machine learning, Ex- plainability, Concept learning, Interpretability. I. INTRODUCTION T IME series extrinsic regression (TSER) refers to to the task of predicting a continuous target variable from an input time series [1]. It appears in many domains—including healthcare (e.g., vital-sign forecasting), finance (e.g., volatil- ity prediction), engineering (e.g., predicting battery state- of-charge or estimating remaining useful life from sensor streams), and environmental monitoring (e.g., pollution esti- mation). In these settings, accurate predictions and trustworthy reasoning are both essential [2]. However, despite recent gains in accuracy, the opacity of modern TSER models remains an obstacle to adoption, particularly when predictions must be understood and validated by domain experts. The demand for interpretability becomes especially difficult to satisfy with current state-of-the-art TSER models. Although recent approaches achieve strong predictive performance, they typically operate as black boxes. Leading approaches include deep neural networks [3], [4], large ensembles [5], and Ran- dOm Convolutional KErnel Transform (ROCKET) models [6]–[8]. Their complexity, large parameter counts, and lack of transparency make it difficult to understand which temporal features drive predictions or how multivariate interactions across variables influence the model’s output. These limita- tions hinder deployment in sensitive or regulated contexts. For example, if a TSER model predicts an impending system failure, engineers need to trace that prediction back to specific sensor behaviors and time intervals, rather than relying on an alert that cannot be meaningfully explained or validated. These interpretability challenges have motivated growing interest in Explainable AI (XAI) research for time series, which aims to bridge the gap between predictive accuracy and actionable insight. While this research direction has made meaningful progress, existing approaches still fall short in several ways. Post-hoc explanation methods, such as saliency maps and feature attribution [9]–[11], attempt to rationalize a model’s decisions after training, but their explanations are often coarse, unstable, or poorly aligned with the model’s true internal reasoning [12]. In contrast, inherently interpretable models embed trans- parency directly into their architecture. Neural additive models (NAMs) [13] and their time-series extension, Neural Addi- tive Time-series Models (NATMs) [14], offer interpretable decompositions but remain

…(Full text truncated)…

This content is AI-processed based on ArXiv data.