Research map / Time series / Foundation models for time series
CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series
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 CauDiTS, a causal disentangled unsupervised domain adaptation framework for multivariate time series that separates domain-common causal rationales from domain-specific correlations and enforces cross-domain causal prototype consistency and intervention invariance.
Problem
Unsupervised domain adaptation for multivariate time series classification under distribution shift, where existing methods extract domain-invariant features via shared extractors and capture spurious shortcut features rather than stable causal patterns among variables.
Approach
An adaptive rationale disentangler uses a learnable binary causal mask over variable interrelation attention to split edges into domain-common causal rationales (used for classification via a causally augmented GCN) and domain-specific non-causal correlations (used for domain discrimination); cross-domain causal prototype consistency (SVD-based prototypes) and backdoor-adjustment-based domain inte
Key results
- CauDiTS achieves average Macro-F1 of 0.7683 on Boiler, outperforming second-best CLUDA (0.680) by 12.99%.
- On WISDM average Macro-F1 improves 21.29% over best baseline (0.607 vs 0.7362 reported as 0.7362 vs 0.607 in table); on HHAR average Macro-F1 0.7191 vs 0.666 for CLUDA (7.97% gain).
- Averaged over all datasets CauDiTS outperforms baselines by 11.81%; ablation shows ARAD adds roughly 8% mean accuracy over L_bs alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CauDiTS instead of CLUDA | UDA for multivariate time series · Boiler, average selected source→target pairs | Macro-F1: 0.680 -> 0.7683 | Helped |
| Use CauDiTS instead of CLUDA | UDA for multivariate time series · WISDM, average selected source→target pairs | Macro-F1: 0.607 -> 0.7362 | Helped |
| Use CauDiTS instead of RAINCOAT | UDA for multivariate time series · HAR, average selected source→target pairs | Macro-F1: 0.905 -> 0.9502 | Helped |
| Use CauDiTS instead of RAINCOAT | UDA for multivariate time series · HHAR, average selected source→target pairs | Macro-F1: 0.646 -> 0.7191 | Helped |
| Add ARAD (ablation) | CauDiTS with Lbs · Boiler, WISDM, HAR, and HHAR | accuracy: Boiler: 0.7832 -> 0.8510; WISDM: 0.6800 -> 0.7396; HAR: 0.7159 -> 0.8462; HHAR: 0.5551 -> 0.6325 | Helped |
| Add Lcon (ablation) | CauDiTS with Lbs + ARAD · Boiler, WISDM, HAR, and HHAR | accuracy: Boiler: 0.8510 -> 0.8812; WISDM: 0.7396 -> 0.7761; HAR: 0.8462 -> 0.9137; HHAR: 0.6325 -> 0.7142 | Helped |
| Add Lcpc (ablation) | CauDiTS with Lbs + ARAD + Lcon · Boiler, WISDM, HAR, and HHAR | accuracy: Boiler: 0.8812 -> 0.8962; WISDM: 0.7761 -> 0.7999; HAR: 0.9137 -> 0.9390; HHAR: 0.7142 -> 0.7539 | Helped |
| Add Ldcls (ablation) | CauDiTS with Lbs + ARAD + Lcon + Lcpc · Boiler, WISDM, HAR, and HHAR | accuracy: Boiler: 0.8962 -> 0.9358; WISDM: 0.7999 -> 0.8375; HAR: 0.9390 -> 0.9731; HHAR: 0.7539 -> 0.8001 | Helped |
| Replace LSTM with RNN or GRU (ablation) | CauDiTS · Selected source→target pairs | Macro-F1 | No clear effect |
| Use CNN instead of TCN or ResNet18 as the backbone (ablation) | CauDiTS · All datasets | Macro-F1 | Helped |
Surprising findings
- Table 1 reports CDAN and other methods outperform some reported numbers inconsistently; the paper notes LSTM is best hidden-state extractor though differences across RNN/GRU are small.
Limitations
- Gains are reported as relative improvements on mean Macro-F1 over few source-target pairs (five pairs per dataset).
- Causal rationale assumptions (per-class unique invariant rationale, observed confounders via random sampling) are assumed rather than verified.
Experimental setup
Datasets: WISDM HAR HHAR Boiler
Benchmarks: WISDM HAR HHAR Boiler
Metrics: Macro-F1 accuracy
Baselines: VRADA CoDATS AdvSKM SASA RAINCOAT CLUDA CAN CDAN DDC DeepCORAL DSAN HoMM
Builds on: DSR iMSDA CLUDA CoTMix SASA
Closest papers by meaning
- Enhancing Multivariate Time-Series Domain Adaptation via Contrastive Frequency Graph Discovery and Language-Guided Adversary Alignment (AAAI 2025 · similarity 0.88)
- Efficient Source-Free Time-Series Adaptation via Parameter Subspace Disentanglement (ICLR 2025 · similarity 0.86)
- Reinforced Cross-Domain Knowledge Distillation on Time Series Data (NeurIPS 2024 · similarity 0.86)
- Boosting Transferability and Discriminability for Time Series Domain Adaptation (NeurIPS 2024 · similarity 0.86)
- Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.86)
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.86)
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.