FLEXtime: Filterbank learning for explaining time series

2024-11-06

arXiv.org

AI Summary

To address the lack of physical interpretability in time-series forecasting models, this paper proposes the first learnable frequency-domain attribution framework tailored for time-series explainability. Methodologically, it introduces differentiable IIR/FIR filter banks to adaptively decompose raw signals into multiple frequency bands; integrates band-wise weighted fusion and gradient-driven importance optimization to enable end-to-end, frequency-aware instance-level attribution—avoiding signal distortion caused by conventional zeroing-based frequency masking. The core contribution lies in pioneering the integration of learnable filter banks into time-series interpretability, establishing a physically grounded, training-stable frequency-domain modeling paradigm. Evaluated on diverse time-series data—including EEG and audio—the framework achieves state-of-the-art average performance across benchmarks. Its attributions are more robust and align better with signal processing priors, thereby supporting trustworthy clinical and engineering decision-making.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Computer Vision: Interpretability, Explainability, and Transparency
• Philosophy and Ethics of AI: Accountability, Interpretability & Explainability

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphs

Abstract

State-of-the-art methods for explaining predictions based on time series are built on learning an instance-wise saliency mask for each time step. However, for many types of time series, the salient information is found in the frequency domain. Adopting existing methods to the frequency domain involves naively zeroing out frequency content in the signals, which goes against established signal processing theory. Therefore, we propose a new method entitled FLEXtime, that uses a filterbank to split the time series into frequency bands and learns the optimal combinations of these bands. FLEXtime avoids the drawbacks of zeroing out frequency bins and is more stable and easier to train compared to the naive method. Our extensive evaluation shows that FLEXtime on average outperforms state-of-the-art explainability methods across a range of datasets. FLEXtime fills an important gap in the time series explainability literature and can provide a valuable tool for a wide range of time series like EEG and audio.

Problem

• Explains time series predictions using frequency band saliency maps.
• Improves interpretability of complex time series data.
• Outperforms state-of-the-art explainability methods across datasets.

Innovation

• Uses bandpass filters for time series decomposition
• Learns optimal frequency band combinations for predictions
• Outperforms state-of-the-art explainability methods