A Targeted Learning Framework for Policy Evaluation with Unobserved Network Interference

TMLR Paper6162 Authors

Abstract: Estimating causal effects under network interference is a fundamental yet challenging task, especially when the network structure is represented as multiple layers or multiple views. In this paper, we consider a heterogeneous network setting, where the ties from different views of the network might achieve varying levels of interference. Meanwhile, dependence among units is allowed, due to information transmission among network ties and latent traits among units sharing ties (i.e., latent dependency). To the best of our knowledge, this setting has not been studied in literature yet. We propose a novel framework that conducts doubly robust estimation on heterogeneous networks with latent dependency. Our approach relies on a new identification strategy and integrates it with targeted maximum likelihood estimation for robust causal effect estimation from observational data. Crucially, our approach remains valid even when the outcome prediction model or data-generating process is misspecified. It also supports counterfactual inference under hypothetical network interventions using only the observed network structure. Experiments on both synthetic and real-world networks show that our approach consistently outperforms existing baselines and can provide robust estimation towards different intervention policies.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Karthikeyan_Shanmugam1

Submission Number: 6162