Discovering Heterogeneous Causal Effects in Relational Data

Real-world and online social and interaction networks are rich sources of human behavioral data. There is growing interest in deriving causal insights from such data which is inherently relational in nature. Causal inference in relational settings has to account for interference, where a unit's outcome may be influenced by the treatments or outcomes of other units. Despite recent progress in causal inference under interference research, there has been limited attention to heterogeneous effects when different units have different responses to treatment and/or a different susceptibility to peer influence, depending on the unit's contexts. This thesis addresses this gap by defining heterogeneous peer influence (HPI) as the general interference that occurs when a unit's outcome may be influenced differently by different peers based on some underlying mechanisms involving their attributes and relationships. Understanding heterogeneity in causal effects facilitates measuring impacts of treatment policies for different subpopulations and uncovering targeted intervention policies custom to relational domains such as viral marketing strategies, risk reduction interventions for infectious diseases, and awareness campaigns for vulnerable groups in social networks. The main objective of this thesis is to develop a framework for expressive causal modeling, sound causal reasoning, and robust estimation of individual, i.e., unit-level, direct and peer effects in the presence of heterogeneous peer influence when the mechanism of influence is unknown. First, I present my initial research focused on developing an observational study design for testing a cause-effect hypothesis using data collected from Twitter, a popular online social network. Using Twitter opinions, I test whether recreational cannabis legalization impacts the development of pro-cannabis attitudes for the population in favor of tobacco vaping, and analyze heterogeneity of causal effects for different states implementing the policy and time since legalization. Second, I discuss my work on Network Structural Causal Model (NSCM) and Network Abstract Ground Graph (NAGG), a framework for expressive causal modeling and sound causal reasoning in networks, along with IDE-Net, an approach for robust individual direct effect estimation when underlying mechanisms of heterogeneous peer influence (HPI) are unknown. Third, I present EgoNetGNN, a novel graph neural network (GNN) architecture, to capture unknown HPI mechanisms that involve not only peer treatments but also attributes of the local neighborhood, including node, edge, and structural attributes for robust heterogeneous peer effect estimation. This thesis integrates causal inference, graph machine learning, and network science to uncover heterogeneous causal effects in complex network settings, paving the way for future multidisciplinary research directions and applications.

Author dc:creator

• Shishir Adhikari (5641526)

• Computer Science

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451422.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451422