Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization

Naiyu Yin*, Hanjing Wang, Yue Yu, Tian Gao, Amit Dhurandhar, Qiang Ji ;

"Identifying low-dimensional, semantic latent causal representations for high-dimensional data has become a dynamic field in computer vision and machine learning. Causal domain generalization methods aim to identify latent causal variables that generate input data and build invariant causal mechanisms for prediction tasks, thereby improving out-of-distribution (OOD) prediction performance. However, there is no consensus on the best approach for selecting causal variables for prediction. Existing methods typically choose causal or anti-causal variables, excluding other invariant, discriminative features. In this paper, we propose using Markov Blanket features due to their property of being the minimal set that possesses the maximum mutual information with the target. To achieve this, we establish a Causal Markov Blanket Representation Learning (CMBRL) framework, which allows for Markov Blanket discovery in the latent space. We then construct an invariant prediction mechanism using the identified Markov Blanket features, making it suitable for predictions across domains. Compared to state-of-the-art domain generalization methods, our approach exhibits robustness and adaptability under distribution shifts."