Estimating treatment effects in networks using domain adversarial learning

Daan Caljon, Jente Van Belle, Wouter Verbeke

Keywords: Interference, Causal Machine Learning, Spillover Effects, Exposure Mapping, Homophily

Abstract:

Estimating heterogeneous treatment effects in networked settings is complicated by interference, meaning that an instance's outcome can be influenced by the treatment status of others. Existing causal machine learning approaches often assume a known exposure mapping
that summarizes how the outcome of a given instance is influenced by
others' treatments, a simplification that is often unrealistic. Furthermore, the interaction between homophily---the tendency of similar instances to connect---and the treatment assignment mechanism has not been explicitly studied before. This interaction can induce a network-level covariate shift, potentially biasing the estimated treatment effects. To address these challenges, we propose HINet---a novel method that integrates Graph Neural Networks (GNNs) with domain adversarial learning. This combination enables learning treatment-invariant node representations, thereby mitigating potential bias caused by covariate shift. Our empirical evaluations on synthetic and semi-synthetic network datasets demonstrate that our approach outperforms existing methods.

Submission Number: 13