TL;DR
This paper introduces HINet, a novel method combining graph neural networks and domain adversarial training to estimate treatment effects in networks without known exposure mappings, effectively addressing covariate shift issues.
Contribution
HINet is the first approach to estimate treatment effects in networks under unknown exposure mappings while mitigating covariate shift through domain adversarial training.
Findings
HINet outperforms existing methods on synthetic datasets.
It effectively reduces bias caused by network-level covariate shift.
Empirical results demonstrate improved accuracy in treatment effect estimation.
Abstract
Estimating heterogeneous treatment effects in network settings is complicated by interference, meaning that the outcome of an instance can be influenced by the treatment status of others. Existing causal machine learning approaches usually assume a known exposure mapping that summarizes how the outcome of a given instance is influenced by others' treatment, a simplification that is often unrealistic. Furthermore, the interaction between homophily -- the tendency of similar instances to connect -- and the treatment assignment mechanism can induce a network-level covariate shift that may lead to inaccurate treatment effect estimates, a phenomenon that has not yet been explicitly studied. To address these challenges, we propose HINet, a novel method that integrates graph neural networks with domain adversarial training. This combination allows estimating treatment effects under unknown…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The proposed method is generally reasonable.
1. The technical contributions are not significant. Both GIN and adversarial training are classical existing approaches. GIN has been widely used in graph learning, and adversarial learning is a standard method for distribution alignment. Exposure mapping learning has also been studied in the literature, as discussed in Section 2. What is the major difference and advantage of the submission compared with existing methods? 2. The key point of the submission is exposure mapping learning. What is
The logical structure of the paper is clear, and the exploration of homophily under network interference is significant.
1. The assumptions in Section 3 still rely on one-hop interference (Markov assumption) and strong ignorability. If higher-order interference exists in real networks, the current Figure 2 and Equation (5) cannot adequately model those dependencies. 2. All results are based on semi-synthetic or fully synthetic datasets; adding a small robustness study on a real-world social graph with unobserved homophily would make the work more complete. 3. As shown in Sections 3 and 5.2, the paper only provides
- The paper addresses an important concern on network homophily in causal inference with interference. The introduction of the concept of network-level covariate shift due to homophily and treatment assignment interaction is a novel perspective. - The proposed HINet framework is methodologically reasonable, combining Graph Neural Networks with counterfactual representation learning to jointly learn exposure effects and estimate treatment effects. - Theoretical analysis and extensive empirical
- The paper lacks rigorous formulation of several key concepts. For instance, the adversarial training framework is not clearly defined or formalized, and the notion of covariate shift in networks is not explicitly characterized. The relationship between network homophily and covariate shift also requires further mathematical clarification to support the proposed intuition. - The algorithmic presentation omits some important implementation details. In particular, the paper does not clearly desc
The experimental design is comprehensive and well-executed, with the authors testing multiple exposure mapping functions including weighted averages, sums, entropy-based measures, and squared weighted averages. The introduction of two new evaluation metrics, PEHNE and CNEE, addresses a genuine gap in the literature regarding how to evaluate treatment effect estimators under interference across multiple counterfactual networks rather than just a single counterfactual scenario.
The most significant theoretical limitation is the lack of guarantees regarding the invertibility of learned representations. As the authors acknowledge, non-invertible representations can lead to biased treatment effect estimates, yet HINet provides no mechanism to ensure or verify invertibility. This is particularly concerning given that domain adversarial training explicitly removes information to achieve balance, potentially making the treatment effect non-identifiable.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
