ShapeMatch: Shapelet-Guided Semi-Supervised Learning for Multivariate Time Series Classification

Xuan-May Le, Ling Luo, Uwe Aickelin, Minh-Tuan Tran

Keywords: Time Series, Semi-Supervised Learning, Transformer, Classification

TL;DR: We propose ShapeMatch, a semi-supervised framework for MTSC. It combines shapelet-guided training and a novel ShapeAug method to improve Deep learning model in low-label settings. ShapeMatch outperforms SOTA models across benchmarks.

Abstract: Multivariate Time Series Classification (MTSC) is crucial for many real-world applications and deep learning models such as Transformer have become the state-of-the-art (SOTA) for MTSC due to their ability to capture complex temporal and spatial dependencies. However, they struggle to perform well without sufficient labelled data, limiting their effectiveness in label-scarce scenarios. Furthermore, the absence of effective augmentation methods for time series data that can enhance generalisation whilst preserving essential temporal structures poses a significant challenge. As a result, despite the success of semi-supervised learning in other domains, these limitations have left its integration with deep learning-based MTSC largely unexplored. To bridge this gap, we propose ShapeMatch, a novel flexible semi-supervised framework designed to enhance MTSC in label-constrained environments. ShapeMatch introduces two key innovations: (1) a hybrid training approach that leverages the classic Shapelet Model to guide the deep learning model in the early stages, capitalising on Shapelets' robustness for low-label regimes, and (2) ShapeAug, a tailored augmentation strategy for multivariate time series that preserves critical structural patterns whilst introducing meaningful variations. Extensive experiments on benchmark datasets demonstrate that ShapeMatch surpasses existing SOTA methods for scenarios with limited labelled data, making it a powerful solution for real-world MTSC applications. Our code is available at http://anonymous.4open.science/r/Shape-Match-MTSC/

Primary Area: learning on time series and dynamical systems

Submission Number: 20302