Research map / Time series / Foundation models for time series
Boosting Transferability and Discriminability for Time Series Domain Adaptation
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 ACON, a time series unsupervised domain adaptation method that combines multi-period frequency features, temporal-frequency mutual distillation, and adversarial alignment in a temporal-frequency correlation subspace.
Problem
Unsupervised domain adaptation for time series classification, where existing methods either ignore frequency features or treat temporal and frequency features equally despite their differing discriminability and transferability.
Approach
Frequency features are more discriminative within a domain while temporal features transfer better across domains; ACON exploits this via bidirectional KL-based mutual learning (frequency→temporal in source, temporal→frequency in target), period-segmented FFT features, and domain adversarial learning on the outer-product temporal-frequency correlation subspace.
Key results
- ACON achieves 74.08% average accuracy on CAP versus 69.13% for the strongest baseline RAINCOAT/DIRT-T (improvement reported as 5.20% over DIRT-T).
- On HHAR-P, ACON reaches 81.74% average accuracy, 10.15% above RAINCOAT (74.21%).
- Ablation on UCIHAR, HHAR-P, WISDM: full model averages 87.85% versus 82.58% without the adversarial/mutual distillation losses LMs, LMt and 65.05% for temporal-only source-only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ACON instead of the strongest reported baseline | Time-series UDA · EMG, FD, PCL, UCIHAR, HHAR-P, WISDM, HHAR-D, CAP | accuracy: 2.75%; 1.30%; 1.31%; 2.74%; 10.15%; 10.70%; 15.85%; 5.20% | Helped |
| add multi-period frequency learning (ablation) | Frequency-only ACON · UCIHAR, HHAR-P, WISDM | accuracy: 66.88 -> 73.47; 51.08 -> 53.16; 56.47 -> 59.10 | Helped |
| add temporal-frequency domain mutual learning (LMs and LMt) (ablation) | ACON with temporal and frequency classification and period learning · UCIHAR, HHAR-P, WISDM | accuracy: 94.05 -> 90.83; 79.49 -> 57.83; 74.19 -> 67.77 | Mixed |
| add domain adversarial learning loss LD (ablation) | ACON with temporal and frequency classification, period learning, and mutual learning · UCIHAR, HHAR-P, WISDM | accuracy: 90.83 -> 97.02; 57.83 -> 81.74; 67.77 -> 84.80 | Helped |
| add frequency classification loss to temporal-only classification (ablation) | ACON · UCIHAR, HHAR-P, WISDM | accuracy: 75.12 -> 94.05; 54.25 -> 79.49; 65.78 -> 74.19 | Mixed |
Surprising findings
- Frequency-domain classification is more discriminative in the source domain but underperforms temporal features in the target domain, showing source discriminability does not imply target discriminability.
- All trade-off hyperparameters are set to 1 without tuning, yet ACON still outperforms baselines; sensitivity analysis shows stable performance.
Limitations
- Method remains unstable on time series with relatively large variances, per the authors.
- Trade-off hyperparameters are fixed to 1 because target-side tuning is unavailable in UDA.
- Evaluation uses randomly sampled source-target pairs (10 per dataset), limiting coverage.
Experimental setup
Datasets: EMG FD PCL UCIHAR HHAR-P WISDM HHAR-D CAP
Benchmarks: EMG FD PCL UCIHAR HHAR-P WISDM HHAR-D CAP
Metrics: accuracy macro-F1
Baselines: Source-only CDAN DeepCoral AdaMatch HoMM DIRT-T CLUDA AdvSKM CoDATS RAINCOAT
Builds on: DANN AdaTime RAINCOAT CoDATS knowledge distillation
Closest papers by meaning
- Breakthrough Sensor-Limited Single View: Towards Implicit Temporal Dynamics for Time Series Domain Adaptation (NeurIPS 2025 · similarity 0.89)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.88)
- Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment (AAAI 2025 · similarity 0.87)
- CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series (ICML 2024 · similarity 0.86)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.85)
- TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts (AAAI 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.