Mask2Cause: End-to-End Causal Discovery via Attention
An innovative framework named Mask2Cause has been introduced by researchers for causal discovery in time series, enabling the direct recovery of causal graphs while forecasting. This approach employs Inverted Variable Embedding along with Adjacency-Constrained Masked Attention, which is trained using either homoscedastic or heteroscedastic objectives to effectively identify causal relationships in both mean and variance. Mask2Cause demonstrates cutting-edge performance on various benchmarks, ranging from synthetic chaotic dynamics to biological simulations, all while minimizing parameter complexity.
Key facts
- Mask2Cause is an end-to-end framework for causal discovery in time series.
- It recovers causal graphs directly during the forecasting forward pass.
- Uses Inverted Variable Embedding and Adjacency-Constrained Masked Attention.
- Trained with homoscedastic or heteroscedastic objectives.
- Captures causal influences in both mean and variance.
- Achieves state-of-the-art causal discovery on diverse benchmarks.
- Reduces parameter complexity compared to standard baselines.
- Tested on synthetic chaotic dynamics and realistic biological simulations.
Entities
Institutions
- arXiv