Title: Estimating Treatment Effects with Differentiated Networked Effects on Graph Data

Abstract:

Deriving individual treatment effects (ITE) from observational graph data is a pivotal component in decision-making processes within sectors such as healthcare and e-commerce. This endeavor is complicated by the presence of interference, a phenomenon where an individual’s outcomes are shaped not only by their own attributes but also by the treatments and covariates of their connected peers. While current methodologies strive to account for this interference to enhance ITE accuracy, they frequently neglect a crucial factor: the differentiated networked effect (DNE). DNE arises from local network structures where neighbors possess varying degrees of significance and scale. Failing to capture DNE leads to a flawed characterization of interference, resulting in inaccurate ITE estimates and potentially erroneous strategic choices. To overcome this limitation, we introduce a novel interference modeling framework featuring a message amplifier alongside two partial attention mechanisms. These attention mechanisms automatically assess the relevance of different neighbors in driving interference, while the message amplifier refines the modeling output by accounting for neighbor scales. Together, these components allow the model to effectively capture DNE. Our evaluations across three real-world graph datasets reveal that our approach surpasses existing techniques for graph-based ITE estimation, thereby validating the necessity of explicitly modeling DNE.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC