Feature Synergy and Interference: An Analysis for Time-Series Classification

Fanzhe Fu, Yang Yang

Keywords: Time Series Classification, Feature-centric Analysis, Feature Fusion, Feature Pruning, ROCKET, InceptionTime, Adaptive Models, Dataset Characteristics

TL;DR: Instead of a universal model for Time Series Classification (TSC), we prove the best strategy (pruning vs. fusion) is predictable from dataset metrics like SNR and entropy. This enables a new, adaptive, feature-centric approach to TSC.

Abstract: The pursuit of a universal, one-size-fits-all model has dominated Time Series Classification (TSC) research. This work challenges that paradigm, arguing that advancing TSC requires a fundamental understanding of feature interplay, not merely more complex architectures. We conduct a series of meticulously designed controlled experiments to dissect the feature spaces of a wide array of representative TSC models, from efficient feature extractors like ROCKET to state-of-the-art deep learning architectures including Transformers and Mamba. For high-dimensional feature extractors, we reveal that the performance bottleneck is dataset-dependent, shifting between feature redundancy and feature noise. We demonstrate that for complex non-linear classifiers, feature pruning can serve as a critical de-noising step on noisy datasets, while for simpler linear models, the full feature set can sometimes be more robust. For a diverse set of nine deep models, we systematically evaluate time-frequency fusion strategies, showing that the optimal choice is intricately linked to both the dataset's intrinsic properties and the model's architectural biases. We uncover clear and widespread evidence of "feature synergy," where fusion provides significant gains, and "feature interference," where it actively degrades performance. Our work pivots the focus from a "model-centric" to a "feature-centric" perspective, providing a new paradigm and a concrete analytical framework for developing adaptive and truly robust TSC solutions.

Primary Area: learning on time series and dynamical systems

Submission Number: 7004