Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains
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 SYNC, a time-aware structural causal model and sequential VAE method that learns static and dynamic causal representations plus drift factors for evolving domain generalization.
Problem
Existing evolving domain generalization (EDG) methods learn data-target correlations over time and are prone to spurious correlations; static causal models ignore dynamic causal factors and drifting causal mechanisms in non-stationary environments.
Approach
A time-aware SCM with static and dynamic causal factors and time-varying drift factors, learned via a sequential VAE with mutual-information disentanglement and supervised contrastive conditional-MI objectives that preserve intra-class compactness across and within domains.
Key results
- SYNC achieves 63.4% worst-case and 73.1% average accuracy overall, the best among compared methods (SDE-EDG overall 55.4 / 71.9 and MMD-LSAE 58.1 / 70.9).
- On RMNIST ablation, full SYNC reaches 45.8 worst / 50.8 average vs 40.5 / 44.1 for the base evolving-pattern-only variant.
- SYNC improves over the best EDG method by 5.3 points in worst-case and 1.2 points in average accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SYNC instead of the strongest causal DG baseline | evolving domain generalization · Overall | worst accuracy: +7.6% | Helped |
| use SYNC instead of the strongest causal DG baseline | evolving domain generalization · Overall | average accuracy: +7.7% | Helped |
| use SYNC instead of the state-of-the-art EDG approach | evolving domain generalization · Overall | worst accuracy: +5.3% | Helped |
| use SYNC instead of the state-of-the-art EDG approach | evolving domain generalization · Overall | average accuracy: +1.2% | Helped |
| add static-dynamic disentanglement loss (ablation) | SYNC on RMNIST · RMNIST | worst accuracy: 40.5 -> 41.9 | Helped |
| add static-dynamic disentanglement loss (ablation) | SYNC on RMNIST · RMNIST | average accuracy: 44.1 -> 45.7 | Helped |
| add static causal loss (ablation) | SYNC on RMNIST · RMNIST | worst accuracy: 41.9 -> 44.1 | Helped |
| add static causal loss (ablation) | SYNC on RMNIST · RMNIST | average accuracy: 45.7 -> 48.7 | Helped |
| add dynamic causal loss (ablation) | SYNC on RMNIST · RMNIST | worst accuracy: 41.9 -> 42.9 | Helped |
| add dynamic causal loss (ablation) | SYNC on RMNIST · RMNIST | average accuracy: 45.7 -> 49.2 | Helped |
| add both static and dynamic causal losses (ablation) | SYNC on RMNIST · RMNIST | worst accuracy: 40.5 -> 45.8 | Helped |
| add both static and dynamic causal losses (ablation) | SYNC on RMNIST · RMNIST | average accuracy: 44.1 -> 50.8 | Helped |
Surprising findings
- Worst-case performance of EDG methods does not significantly exceed that of static DG methods, despite their advantage in average accuracy.
Limitations
- Relies on the entropy-inequality conditions of Proposition 1 holding, which are argued to be easy to satisfy but not verified empirically.
- Evaluated only on relatively small benchmark datasets built on DomainBed with fixed small MLP/CNN architectures.
Experimental setup
Datasets: Circle Sine RMNIST Portraits Caltran PowerSupply ONP
Benchmarks: Circle Sine RMNIST Portraits Caltran PowerSupply ONP
Metrics: worst-case accuracy average accuracy
Baselines: ERM Mixup MMD MLDG RSC MTL FISH CORAL AndMask DIVA IRM IIB
Builds on: LSSAE DomainBed IRM Gumbel-Softmax
Closest papers by meaning
- Enhancing Evolving Domain Generalization through Dynamic Latent Representations (AAAI 2024 · similarity 0.90)
- UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems (NeurIPS 2025 · similarity 0.88)
- Generalizing across Temporal Domains with Koopman Operators (AAAI 2024 · similarity 0.88)
- A Fixed-Point Approach for Causal Generative Modeling (ICML 2024 · similarity 0.88)
- Learning Structural Causal Models from Ordering: Identifiable Flow Models (AAAI 2025 · similarity 0.87)
- Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations (ICLR 2025 · similarity 0.87)
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.