TFEC: Multivariate Time-Series Clustering via Temporal-Frequency Enhanced Contrastive Learning

TL;DR

The paper addresses multivariate time-series clustering with contrastive learning, identifying two key limitations: lack of clustering-aware sample construction and distortion from augmentations. It introduces TFEC, combining Temporal-Frequency Co-EnHancement (CoEH) with a synergistic dual-path architecture (PGCL and READ) to jointly optimize representation fidelity and cluster structure. A joint loss $\mathcal{L}_{\text{total}} = \beta \mathcal{L}_{\text{con}} + (1-\beta)\mathcal{L}_{\text{recon}}$ balances cluster-discriminative signals with reconstruction stability, enabling end-to-end training. Empirically, TFEC achieves an average NMI gain of 4.48% over SOTA across six UEA datasets, and ablations validate the contribution of each component. Code is available at https://github.com/yueliangy/TFEC.

Abstract

Multivariate Time-Series (MTS) clustering is crucial for signal processing and data analysis. Although deep learning approaches, particularly those leveraging Contrastive Learning (CL), are prominent for MTS representation, existing CL-based models face two key limitations: 1) neglecting clustering information during positive/negative sample pair construction, and 2) introducing unreasonable inductive biases, e.g., destroying time dependence and periodicity through augmentation strategies, compromising representation quality. This paper, therefore, proposes a Temporal-Frequency Enhanced Contrastive (TFEC) learning framework. To preserve temporal structure while generating low-distortion representations, a temporal-frequency Co-EnHancement (CoEH) mechanism is introduced. Accordingly, a synergistic dual-path representation and cluster distribution learning framework is designed to jointly optimize cluster structure and representation fidelity. Experiments on six real-world benchmark datasets demonstrate TFEC's superiority, achieving 4.48% average NMI gains over SOTA methods, with ablation studies validating the design. The code of the paper is available at: https://github.com/yueliangy/TFEC.

Figures (2)

• Figure 1: Overall framework of TFEC. CoEH generates low-distortion EME. The dual-path architecture processes EME: the PGCL path performs pseudo-label guided contrastive learning on cluster structures based on high-confidence samples, while the READ path stabilizes representations via mask reconstruction.
• Figure 2: Clustering performance comparison on six UEA datasets using ACC, F1 and NMI as evaluation metrics.