Research map / Time series / Foundation models for time series
Cross-Domain Contrastive Learning for Time Series Clustering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CDCC, an end-to-end deep time series clustering model that jointly learns temporal (BiLSTM) and frequency (FFT + conv) encoders with intra- and cross-domain instance- and cluster-level contrastive losses, evaluated on 40 UCR datasets.
Problem
Deep time series clustering usually separates representation learning from clustering, so clustering loss cannot guide feature extraction, and most methods ignore frequency-domain information.
Approach
Jointly optimizes clustering and representation in an end-to-end manner using instance-level and cluster-level contrastive losses within the temporal and frequency domains, plus cross-domain contrastive constraints applied only to augmented samples, with cluster assignments taken from the temporal domain.
Key results
- CDCC achieves the highest average NMI (0.601) and average RI (0.877) across 40 UCR datasets, with best NMI on 18 and best RI on 26 datasets.
- Average rank 2.075 for NMI and 1.800 for RI, best among the compared methods (TSTCC, TST, FeatTS, STCN, R-Clust, TCGAN).
- Ablations show removing cluster-level cross-domain contrast causes a notable drop, and removing both cross-domain and frequency contrast causes a significant decline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDCC instead of R-Clust | Time series clustering · UCR 40 datasets | Average NMI: 0.524 -> 0.601 | Helped |
| Use CDCC instead of R-Clust | Time series clustering · UCR 40 datasets | Average RI: 0.823 -> 0.877 | Helped |
| Remove instance-level cross-domain contrast (ablation) | CDCC · UCR 40 datasets | Clustering performance (NMI and RI) | Hurt |
| Remove cluster-level cross-domain contrast (ablation) | CDCC · UCR 40 datasets | Clustering performance (NMI and RI) | Hurt |
| Remove frequency-domain contrast (ablation) | CDCC · UCR 40 datasets | Clustering performance (NMI and RI) | Hurt |
| Remove frequency-domain and cross-domain contrast (ablation) | CDCC · UCR 40 datasets | Clustering performance (NMI and RI) | Hurt |
| Increase frequency augmentation perturbation rates θ/ϵ (ablation) | CDCC · ArrowHead, CBF, Fungi, SwedishLeaf | NMI and RI | Mixed |
| Vary the intra-domain/cross-domain balancing coefficient λ (ablation) | CDCC · ArrowHead, CBF, Fungi, SwedishLeaf | NMI and RI | Mixed |
Surprising findings
- Applying the cross-domain constraint to original (non-augmented) data caused overfitting, so it is restricted to augmented samples.
Limitations
- Designed mainly for periodic data; aperiodic data (e.g., image, device) requires further exploration.
- Clustering outputs are taken from the temporal domain, so frequency-domain information is used only indirectly for labeling.
Experimental setup
Datasets: UCR Time Series Archive (40 datasets)
Benchmarks: UCR time series archive (40 datasets)
Metrics: NMI Rand Index average rank Wilcoxon signed-rank test Nemenyi post-hoc test
Baselines: TSTCC TST FeatTS STCN R-Clust TCGAN
Builds on: contrastive clustering (Li et al. 2021) InfoNCE TS-TCC time-frequency consistency (Zhang et al. 2022)
Closest papers by meaning
- GeCC: Generalized Contrastive Clustering with Domain Shifts Modeling (AAAI 2025 · similarity 0.88)
- Parametric Augmentation for Time Series Contrastive Learning (ICLR 2024 · similarity 0.85)
- Soft Contrastive Learning for Time Series (ICLR 2024 · similarity 0.84)
- Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment (AAAI 2025 · similarity 0.84)
- Deep Contrastive Graph Learning with Clustering-Oriented Guidance (AAAI 2024 · similarity 0.84)
- Super Deep Contrastive Information Bottleneck for Multi-modal Clustering (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.