Abstract

Most existing causal discovery algorithms rely on the assumptions that data are independent and identically distributed (i.i.d.). However, many real-world domains, such as biological and social networks, violate the i.i.d. assumption and consist of interacting entities whose attributes exhibit complex relational and causal dependencies, breaking the SUTVA assumption and leading to interference. To address these challenges and to facilitate causal reasoning in network settings, I will present two recent contributions that develop graphical models and algorithms for causal discovery in network data in the presence of cycles and latent variables. The first contribution introduces relational acyclification, an operation specifically designed for cyclic relational causal models that enables formal analysis of identifiability in cyclic relational structures. Under the assumptions of relational acyclification and σ-faithfulness, we establish that the Relational Causal Discovery (RCD) algorithm (Maier et al., 2013) is sound and complete for models containing cyclic dependencies. The second contribution presents RelFCI, a causal discovery algorithm that is sound and complete for relational data subject to latent confounding. In this work, we further derive soundness and completeness guarantees for relational d-separation in the presence of latent variables, thereby extending causal discovery theory to a broader class of relational systems.