Research map / Time series / Foundation models for time series
Graph-Aware Contrasting for Multivariate Time-Series Classification
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 MEGA, a self-supervised contrastive framework for multivariate time-series classification that adds graph augmentations (node and edge), node- and graph-level contrasting, and multi-window temporal contrasting to capture spatial and temporal consistency.
Problem
Contrastive learning for multivariate time-series (MTS) classification mainly enforces temporal consistency and neglects spatial consistency, i.e., the stability of individual sensors and their inter-sensor relationships.
Approach
Graph-based contrastive learning tailored to MTS: sensors form graphs per window, augmented by frequency (DWT noise) and temporal permutation at node level and by randomizing weak sensor correlations at edge level; GNN encoders feed node-level (sensor) and graph-level (batch sample) contrastive losses plus a multi-window predictive temporal loss.
Key results
- MEGA achieves best accuracy on four of six datasets; on HAR it reaches 94.27% accuracy vs 92.78% for the best baseline (TS2Vec), and on ISRUC 84.22% vs 80.50% (TS-TCC).
- On FM and EW, MEGA is second best (52.00% and 61.62% accuracy), behind MHCCL (52.40%) and SimCLR (62.05%).
- Ablation on HAR and ISRUC shows removing node-level or graph-level contrasting lowers accuracy by about 2% on HAR, and removing MWTC drops ISRUC accuracy by 5.9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MEGA instead of the strongest reported baseline | MTS classification · HAR | accuracy: 92.78$\pm$0.32 -> 94.27$\pm$0.12 | Helped |
| Use MEGA instead of the strongest reported baseline | MTS classification · ISRUC | accuracy: 80.50$\pm$0.42 -> 84.22$\pm$0.17 | Helped |
| Use MEGA instead of the strongest reported baseline | MTS classification · FM | accuracy: 52.40$\pm$2.28 -> 52.00$\pm$1.54 | Hurt |
| Use MEGA instead of the strongest reported baseline | MTS classification · EW | accuracy: 62.05$\pm$0.85 -> 61.62$\pm$0.30 | Hurt |
| Add node augmentation (ablation) | MEGA · HAR | accuracy: 92.97$\pm$0.23 -> 94.27$\pm$0.12 | Helped |
| Add node augmentation (ablation) | MEGA · ISRUC | accuracy: 83.86$\pm$0.20 -> 84.22$\pm$0.17 | Helped |
| Add edge augmentation (ablation) | MEGA · HAR | accuracy: 93.67$\pm$0.11 -> 94.27$\pm$0.12 | Helped |
| Add edge augmentation (ablation) | MEGA · ISRUC | accuracy: 83.87$\pm$0.18 -> 84.22$\pm$0.17 | Helped |
| Add multi-window temporal contrasting (MWTC) (ablation) | MEGA · HAR | accuracy: 93.60$\pm$0.19 -> 94.27$\pm$0.12 | Helped |
| Add multi-window temporal contrasting (MWTC) (ablation) | MEGA · ISRUC | accuracy: 78.32$\pm$0.45 -> 84.22$\pm$0.17 | Helped |
| Add graph-level contrasting (GC) (ablation) | MEGA · HAR | accuracy: 92.10$\pm$0.09 -> 94.27$\pm$0.12 | Helped |
| Add node-level contrasting (NC) (ablation) | MEGA · ISRUC | accuracy: 81.29$\pm$0.12 -> 84.22$\pm$0.17 | Helped |
Surprising findings
- Removing MWTC reduces HAR accuracy by only 0.67% but ISRUC accuracy by 5.9%, showing dataset-dependent contribution of temporal contrasting.
Limitations
- Window size analysis suggests small windows are better, and results depend on hyperparameters such as the number of retained edges per view.
- Does not achieve best results on FM and EW datasets, where it is second best.
Experimental setup
Datasets: HAR ISRUC ArticularyWordRecognition FingerMovements EigenWorms SpokenArabicDigitsEq
Benchmarks: HAR ISRUC AWR FM EW SAD
Metrics: accuracy macro-averaged F1
Baselines: SimCLR CPC TNC TS-TCC TS2Vec MHCCL
Builds on: SimCLR CPC TS2Vec TS-TCC MHCCL
Closest papers by meaning
- Role Hypergraph Contrastive Learning for Multivariate Time-Series Analysis (AAAI 2026 · similarity 0.89)
- Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data (AAAI 2024 · similarity 0.87)
- Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection (ICML 2025 · similarity 0.86)
- Parametric Augmentation for Time Series Contrastive Learning (ICLR 2024 · similarity 0.85)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.85)
- Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering (NeurIPS 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.