Bounding Spill-over Effect under Structural Uncertainty

Abstract: Causal inference on graphs has attracted increasing attention due to possible interactions among units. One main challenge is the spill-over effects, i.e., the influence of treatments on neighboring nodes on the target outcome. However, the observed graphs may suffer from inconsistency between local network structure and the interference mechanism, which invalidates the identifiability guaranty of existing spill-over estimators. To address the challenge, we propose learning to bound spill-over effects under local structural uncertainty, which aims to obtain optimization-based bounds for uncertain nodes over a learned feasible set of probably consistent ego-graphs. Specifically, we start by introducing a structure proposal network that maps the ego-graph of uncertain nodes to probably consistent candidate graphs, where a spill-over effect estimator is introduced to explore the upper and lower limits of the spill-over effect by traversing the feasible ego-graph space defined by the generator. The generated ego-graphs are constrained to preserve the incomplete information (\textbf{closeness}) and be indistinguishable from the consistent ego-graphs sampled from stable nodes (\textbf{consistency}), whereas the spill-over estimator is constrained to be compatible with the observed outcomes on the network (\textbf{faithfulness}). We formulate the above objectives as a constrained, bi-level adversarial learning framework, where an efficient and stable EM-based objective is proposed to solve the optimization problem. Experiments on both simulated and semi-simulated datasets show the effectiveness of the proposed method.

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

Assigned Action Editor: ~Lu_Zhang3

Submission Number: 7160