Original title: When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate
Florent Forest, Anlei Wei, Olga Fink · September 2026
Predicting numbers from data tracked over time is common in healthcare and finance. Top AI models make good predictions, but they are black boxes. People cannot see why the model made a choice, and tools that explain them later are often noisy or unstable. Researchers created a new model called MAGNETS to solve this. Instead of hiding its choices, MAGNETS learns a small set of clear concepts straight from the data without human labels. It uses learned masks to select specific data points and combine them. This shows exactly which features drive a prediction and the exact times they matter. Finally, the model connects these concepts in a clear, visible way. Tests on real and synthetic data showed MAGNETS matches the accuracy of black-box models while beating older interpretableYou can look at the model's reasoning and understand why it made a specific prediction, not just get a number. systems, especially when multiple features interact.
Imagine a teacher grading a student's essay. Instead of just giving a score, the teacher marks specific sentences and words that influenced the grade—like 'strong thesis here' or 'weak evidence here'. Now suppose the teacher had to do this for every essay without being told what to look for, and the marks had to be consistent across all essays. In this paper, the model learns to place its own 'marks' on time series data, highlighting which moments and measurements matter, and then combines those highlights into a final prediction.
When a hospital uses AI to predict patient deterioration from vital signs, a black-box model is a liability—you cannot trust what you cannot inspect. MAGNETS lets a doctor see not just the prediction, but which hours of data and which features drove it, without needing a separate explanation tool that might be wrong. This makes high-stakes time series predictions safer to act on in healthcare, finance, and engineering.
The abstract does not report results on very long time series or high-dimensional data, so scalability to such settings is unclear. It also does not compare against post-hoc explanation methods on explanation faithfulness metrics, only against interpretable baselines. The approach is demonstrated on regression tasks; classification is not addressed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate — Florent Forest, Anlei Wei, Olga Fink — Data Mining and Knowledge Discovery — doi:10.1007/s10618-026-01267-y
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.