Continuous causal structure learning from incremental instances and feature spaces

Abstract

Learning causal structure in a complex system is crucial for causal inference and decision-making. Yet it remains a challenging problem due to the presence of unknown and varying feature spaces in most real-world scenes. Existing methods, including constraint-based and score-based methods, mainly rely on strict constraints, e.g., faithfulness, and cannot handle varying feature spaces. We propose a new gradient-based method (C2SLIFS) for learning causal structures from dynamic observational data with incremental instances and feature spaces. The C2SLIFS is designed as a nonlinear model with hidden variables, and utilizes the strategies of continuously optimizing reconstruction loss using dynamic weighting coefficients and equality constraints to update causal relationships, ensure acyclicity and avoid faithfulness constraints. We demonstrate the competitive performance of the proposed method through empirical evaluations against existing state-of-the-art methods. Besides that, two real-world case studies on two Bayesian network datasets (BNs), i.e., a protein signal network (Sachs) and a medical diagnostic alarm message network (Alarm), are conducted to elaborate on the effectiveness of C2SLIFS in real-world scenes. Code implementing the proposed algorithm is open-source and publicly available at https://github.com/youdianlong/C2SLIFS.