Research map / Time series / Foundation models for time series
Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment
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 ConFGD, a contrastive frequency graph discovery framework for multivariate time-series unsupervised domain adaptation, plus a language-guided adversary alignment module (ConFGD+) that aligns classification features with GPT-3 label embeddings.
Problem
Unsupervised domain adaptation for multivariate time-series classification, where existing methods align only intra-series temporal features and ignore inter-series frequency dependencies and external label knowledge.
Approach
Decompose series with DWT into frequency channels, learn a frequency correlation graph via a graph discovery encoder/decoder aligned across domains, add momentum-contrastive frequency-contextual learning, and align classification features to precomputed LLM label-text embeddings (LAA module, usable with other UDA models).
Key results
- ConFGD reaches 94.86% accuracy on HAR vs 91.21% for the best baseline CLUDA; ConFGD+ reaches 96.81%.
- On WISDM, ConFGD achieves 79.15% accuracy vs 72.32% for CLUDA (macro F1 62.25 vs 53.92).
- On HHAR, ConFGD+ reaches 84.11% accuracy vs 75.90% for CLUDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FGD and FCCL to the UDA framework | ConFGD · HAR | Avg. Acc (%): 91.21 -> 94.86 | Helped |
| Add FGD and FCCL to the UDA framework | ConFGD · WISDM | Avg. Acc (%): 72.32 -> 79.15 | Helped |
| Add FGD and FCCL to the UDA framework | ConFGD · HHAR | Avg. Acc (%): 75.90 -> 83.08 | Helped |
| Add LAA | ConFGD · HAR | Avg. Acc (%): 94.86 -> 96.81 | Helped |
| Add LAA | ConFGD · WISDM | Avg. F1 (%): 62.25 -> 65.42 | Helped |
| Add LAA | ConFGD · HHAR | Avg. Acc (%): 83.08 -> 84.11 | Helped |
| Add LAA | CLUDA · WISDM | Macro-F1: +15.50% | Helped |
| Remove frequency graph losses LF C and LF F (ablation) | ConFGD · HAR and WISDM | Accuracy and Macro-F1 | Hurt |
| Remove frequency-contextual contrastive loss LCL (ablation) | ConFGD · HAR and WISDM | Accuracy and Macro-F1 | Hurt |
| Remove LCL, LF C, and LF F (ablation) | ConFGD · HAR and WISDM | Accuracy and Macro-F1 | Hurt |
| Remove domain loss Ldomain (ablation) | ConFGD · HAR and WISDM | Accuracy and Macro-F1 | Hurt |
| Remove positional encoding (ablation) | ConFGD · HAR and WISDM | Accuracy and Macro-F1 | Hurt |
Surprising findings
- LLaMA2 embeddings give higher F1 while GPT-3 gives slightly higher accuracy in the LAA evaluation, so embedding source choice shows mixed effects.
Limitations
- Evaluated only on three HAR/WISDM/HHAR activity datasets with source-target pairs defined by participants.
- LAA relies on dataset label names as prompts, so benefit depends on meaningful label text.
Experimental setup
Datasets: HAR WISDM HHAR
Benchmarks: HAR WISDM HHAR
Metrics: accuracy macro F1 standard deviation
Baselines: VRADA CoDATS AdvSKM CAN CDAN DDC DeepCORAL DSAN HoMM MMDA CLUDA TCN
Builds on: CLUDA MoCo DANN VCDN SASA
Closest papers by meaning
- CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series (ICML 2024 · similarity 0.88)
- Boosting Transferability and Discriminability for Time Series Domain Adaptation (NeurIPS 2024 · similarity 0.87)
- Reinforced Cross-Domain Knowledge Distillation on Time Series Data (NeurIPS 2024 · similarity 0.86)
- TransPL: VQ-Code Transition Matrices for Pseudo-Labeling of Time Series Unsupervised Domain Adaptation (ICML 2025 · similarity 0.85)
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.85)
- COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.85)
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.