A Unified View of Causal and Non-causal Feature Selection

In this article, we aim to develop a unified view of causal and non-causal feature selection methods. The will fill in the gap research relation between two types Based on Bayesian network framework information theory, first show that methods share same objective. That is find Markov blanket class attribute, theoretically optimal set for classification. We then examine assumptions made by when searching set, unify mapping them restrictions structure model studied problem. further analyze detail how structural lead different levels approximations employed their search, which result sets found with respect set. With view, can interpret output from perspective derive error bounds both Finally, present practical understanding using extensive experiments synthetic data various real-world data.