When, How Long and How Much? Interpretable Neural Networks for Time Series Regression by Learning to Mask and Aggregate
MAGNETS is a neural architecture for time series regression that learns input-dependent masks to identify relevant temporal regions and aggregates these into interpretable concepts without needing concept supervision. It decomposes prediction into masking, aggregation, concept bottlenecking, and prediction layers, achieving accuracy competitive with black-box models while providing transparent explanations.