Data-Adaptive Dynamic Time Warping-Based Multivariate Time Series Fuzzy Clustering

Abstract

Multivariate time series (MTS) clustering has become a critical research area. Current methods typically rely on space projection or representation learning for clustering but tend to overlook the significance and contribution of MTS dimensions, leading to a failure in accurately modeling the intricate correlations and dependencies among dimensions. Meanwhile, the lack of adaptive regulation for MTS dimensions in distance measures significantly impacts clustering accuracy. In view of these issues, we propose a data-adaptive dynamic time warping (DTW) based fuzzy clustering method for MTS. This method utilizes locally weighted DTW as the kernel distance measure, enabling the adaptive regulation of MTS dimensions. To address the non-convex optimization problem associated with DTW-based clustering, we formulate a comprehensive objective function and present an efficient optimization method based on closed-form solutions. This unsupervised learning method significantly improves the precision of DTW, leading to more accurate and interpretable clustering outcomes. Extensive experiments conducted on eight public datasets, along with comparisons to 10 benchmark methods, demonstrate the competitive performance of our method in terms of both accuracy and efficiency.

Keywords

• Machine learning
• dynamic time warping
• multivariate time series
• fuzzy clustering