The task the researchers tackle, known as time series extrinsic regression, involves predicting a continuous target variable from an input time series. It underpins vital-sign forecasting in healthcare, volatility prediction in finance, pollution estimation in environmental monitoring, and remaining-useful-life estimation in engineering. In all of these settings, accuracy alone is not enough. If a model warns of an impending system failure, engineers need to trace that warning back to specific sensor behaviors and specific time intervals. Yet the strongest current models, from deep convolutional networks to ROCKET-style random kernel transforms and large ensembles, achieve their performance precisely by hiding their reasoning in millions of parameters.

The dominant workaround has been post-hoc explanation: training an opaque model first, then applying saliency maps or feature attribution methods such as Integrated Gradients or DeepLIFT to rationalize its decisions afterwards. The EPFL team argues that this approach is fundamentally fragile. Post-hoc explanations are often coarse, noisy, or unstable, and they are not guaranteed to reflect the model’s true internal reasoning. Different attribution methods can disagree with one another, and because interpretability constraints are imposed only after training, the explanations may bear little relation to how the prediction was actually computed. Meanwhile, inherently interpretable alternatives such as Neural Additive Models and their time-series extension NATMs decompose predictions into per-feature contributions but remain fundamentally univariate, unable to capture interactions between variables. Concept-based models offer richer structure but typically require concept annotations that rarely exist for time series data.

MAGNETS resolves this impasse by embedding transparency directly into the architecture while still allowing a powerful neural network to do the heavy lifting where it cannot corrupt interpretability. The pipeline unfolds in four stages. First, a mask generation network, implemented as a one-dimensional U-Net, examines the input time series and proposes a set of binary masks, one per channel and mask index, that select the temporal regions relevant to the prediction. Second, each mask is applied to the raw input by element-wise multiplication, zeroing out everything outside the selected regions. Third, the masked values are aggregated over time into a single scalar per channel-mask pair, using a simple summation that captures both the duration of the relevant pattern and its intensity. Finally, these aggregated features pass through a linear concept bottleneck and a strictly linear prediction layer, so every output can be traced back, step by transparent step, to specific parts of the original signal.

The technical heart of the method lies in how the binary masks are trained. Because hard, all-or-nothing masks would normally block gradient flow, the authors employ the Straight-Through Gumbel-Softmax estimator: during the forward pass the masks are binarized so that the selected regions are crisp and unambiguous, while during the backward pass gradients flow through a relaxed, differentiable version. The team deliberately chose binary rather than soft masks, arguing that in continuous time series, where the precise magnitude and shape of values matter, blurred probabilistic masks would obscure exactly which parts of the signal drive the prediction. The masks are also input-specific rather than global, meaning the model can point to different relevant intervals for different individual samples, and a single mask may select several disjoint intervals within a channel.

Two regularization terms keep the learned concepts compact and non-redundant. A sparsity penalty, based on the L1 norm of the bottleneck weights, encourages each concept to depend on only a small subset of the aggregated features. An orthogonality loss, computed as the squared Frobenius norm of the deviation of the weight Gram matrix from the identity, discourages different concepts from encoding similar combinations of features. Together, these constraints produce concepts that are both sparse and distinct, making it clear which masked temporal regions and which channels drive each concept activation. The final prediction is a weighted linear combination of the concept activations, so each concept contributes additively and with an explicit, inspectable weight.

To test whether the model actually recovers the right reasoning, the researchers built four synthetic datasets with known ground-truth logic, ranging from a simple univariate rule, where the target equals the area under the curve above a threshold, to a challenging trivariate task where the target is a weighted sum of three conditional areas defined by pairwise comparisons between channels. The results were striking. MAGNETS outperformed all black-box baselines, often by a substantial margin, on the Bivariate and Trivariate-2 tasks, and beat every existing interpretable model across all multivariate settings. On the hardest trivariate dataset, the regularized bottleneck learned three disentangled concepts that corresponded directly to the correct channel relationships, including one concept that depended jointly on two channels, reconstructing the underlying symbolic logic without any supervision. By contrast, DeepLIFT attributions for a black-box CNN fluctuated sharply across neighboring time steps and failed to capture the necessary multivariate relations.

The evaluation extended to ten real-world datasets spanning household power consumption, wind turbine power, benzene concentration, Beijing air quality, flood modeling, battery degradation, and a newly released bridge degradation benchmark. Here, unconstrained black-box models generally retained the lowest error, as expected, but MAGNETS narrowed the gap dramatically and even surpassed all baselines on FloodModeling2 and BatteryDegradation1. Among interpretable models, it was the consistent leader, achieving the lowest root mean square error on seven of the ten datasets while remaining highly competitive on the rest. Notably, the team applied a single fixed configuration across all tasks, with three concepts and ten masks per channel, and a sensitivity analysis confirmed that performance remained stable across a wide range of hyperparameters, demonstrating that the architecture does not depend on per-dataset tuning.

The qualitative analyses may prove the most compelling for practitioners. On the bridge degradation dataset, MAGNETS discovered a meaningful dependency between bridge displacement and train load, consistently isolating intervals where train load exceeded roughly 2.0 and highlighting the corresponding displacement values. This supports a concrete, testable hypothesis: that structural degradation depends primarily on displacement during intervals of critical loading. On the battery dataset, the model focused on the initial voltage drop of a constant-current discharge, a known indicator of internal resistance growth, and the tail of an open-circuit voltage discharge, which reflects remaining active material capacity. These physically meaningful patterns emerged automatically, aligned with battery chemistry, and without any concept annotations.

The work directly confronts the long-standing performance-interpretability trade-off, showing that for regression tasks driven by discrete, localized temporal events, transparency need not come at the cost of accuracy, and can even enhance it. The authors acknowledge limitations that open future directions: real-world explanation correctness still lacks quantitative ground truth, the current aggregation is restricted to summation, and extending the framework to time series classification is a natural next step. But the core message is clear and likely to resonate far beyond machine learning research. By learning to mask and aggregate, MAGNETS answers the three questions its title poses, when a pattern occurs, how long it lasts, and how much it contributes, in a form that domain experts can actually inspect, validate, and act upon. In safety-critical domains from healthcare to civil infrastructure, that may matter more than another decimal point of accuracy.

Subject of Research: Inherently interpretable neural networks for time series extrinsic regression using learned mask-based concept aggregation

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

Article References: Forest, F., Wei, A., & Fink, O. (2026). When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate. Data Mining and Knowledge Discovery, 40(6), Article 103. https://doi.org/10.1007/s10618-026-01267-y

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01267-y

Keywords: time series regression, interpretable machine learning, explainable AI, concept bottleneck models, neural networks, mask learning, EPFL, battery degradation, structural health monitoring, Data Mining and Knowledge Discovery, U-Net, Gumbel-Softmax