Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment

ICLR 2026 Conference SubmissionAnonymous Authors
Dynamic NetworkInterferenceCausality
Abstract:

In recent years, estimating causal effects of treatment on the outcome variable in network environments has attracted growing interest. The intrinsic interconnectedness of network and the attendant violation of the SUTVA assumption have prompted a wave of treatment effect estimation methods tailored to network settings, yielding considerable progress such as capturing hidden confounders by leveraging auxiliary network structure. Nevertheless, despite these advances, the existing methods: (i) mainly focus on the static network, overlooking the dynamic nature of many real-world networks and confounders that evolve over time; (ii) assume the absence of dynamic network interference where one unit’s treatment can affect its neighbors’ outcomes. To address these two limitations, we first define a new estimand of treatment effects accounting for interference in a dynamic network environment, i.e., CATE-ID, and establish its identifiability under such an environment. Then we accordingly propose DSPNET, a framework tailored specifically for treatment effect estimation in dynamic network environment, that leverages historical information and network structure to capture time-varying confounders and model dynamic interference. Extensive experiments demonstrate the superiority of our proposed method compared to state-of-the-art approaches.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces CATE-ID, a new estimand for treatment effects under dynamic network interference, and proposes DSPNET, a framework leveraging historical information and network structure to capture time-varying confounders. It resides in the 'Time-Varying Confounders and Dynamic Interference' leaf under 'Temporal Dynamics and Network Evolution', which contains four papers total. This leaf addresses evolving confounders and interference patterns where treatment effects change over time, distinguishing it from static network methods. The relatively small cluster suggests this is an active but not overcrowded research direction within the broader taxonomy of fifty papers.

The taxonomy reveals neighboring leaves addressing related temporal challenges: 'Temporal and Spatial Spillover Modeling' captures both carryover and spatial effects in panel data, while 'Network Change and Rewiring Effects' focuses on treatment-induced structural changes. The paper's emphasis on dynamic interference and time-varying confounders positions it at the intersection of these concerns, diverging from purely spatial spillover models and from methods assuming fixed network topology. Broader branches like 'Methodological Frameworks for Interference Modeling' provide foundational identification theory, while 'Causal Discovery and Graph Learning' tackles structure learning—domains the paper builds upon but does not directly contribute to.

Among thirty candidates examined, the analysis found limited prior work overlap. The CATE-ID estimand and identifiability contribution examined ten candidates with none clearly refuting it, suggesting theoretical novelty in formalizing this estimand for dynamic settings. The DSPNET framework similarly showed no refutations across ten candidates. However, the interference representation learning via environment exposure contribution encountered one refutable candidate among ten examined, indicating some overlap in how prior work models interference through environmental features. The search scope was constrained to top-K semantic matches plus citation expansion, not an exhaustive field survey.

Based on the limited search of thirty candidates, the work appears to occupy a relatively novel position within its immediate research cluster, particularly in defining the CATE-ID estimand and establishing identifiability for dynamic networks. The single refutation for the interference representation component suggests incremental overlap in modeling techniques, though the overall framework integrates these elements in a context-specific manner. The analysis does not cover the full breadth of dynamic network literature, and deeper examination of longitudinal causal inference methods outside the top-K matches may reveal additional connections.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: treatment effect estimation in dynamic networks with interference. This field addresses the challenge of estimating causal effects when units are embedded in networks where both the network structure and the interference patterns evolve over time. The taxonomy reveals several major branches: Methodological Frameworks for Interference Modeling develops foundational techniques for handling spillover effects; Temporal Dynamics and Network Evolution focuses on time-varying confounders and dynamic treatment regimes; Specialized Network Contexts and Applications examines domain-specific settings such as peer effects in social networks (Peer effects in networks[3]) and bipartite structures (Bipartite causal inference with[9]); Causal Discovery and Graph Learning for Networks tackles the problem of uncovering causal relationships from observational network data (Causalgnn[6], Causal message-passing for experiments[7]); Spatiotemporal Forecasting with Causal Mechanisms integrates predictive modeling with causal reasoning (Deciphering spatio-temporal graph forecasting[17]); and Network Dynamics and Temporal Causality explores how causal structures themselves change over time. These branches collectively address the interplay between network topology, temporal evolution, and interference. A particularly active line of work centers on dynamic treatment regimes under interference, where treatment decisions must adapt to evolving network conditions and time-varying confounders. Modeling Interference for Treatment[0] sits squarely within the Temporal Dynamics and Network Evolution branch, specifically addressing time-varying confounders and dynamic interference. It shares thematic ground with Dynamic treatment regimes with[5] and Dynamic Treatment Regimes on[41], which similarly tackle sequential decision-making in networked settings, and with Estimating dynamic treatment regimes[43], which also emphasizes adaptive strategies. Compared to these neighbors, Modeling Interference for Treatment[0] places particular emphasis on the interplay between temporal confounder adjustment and interference patterns, whereas Dynamic treatment regimes with[5] may focus more on longitudinal treatment sequences and Estimating dynamic treatment regimes[43] on estimation robustness. Open questions across this cluster include scalability to large networks, identification under partial observability, and the integration of causal discovery with treatment effect estimation.

Claimed Contributions

CATE-ID estimand and identifiability for dynamic networks with interference

The authors introduce CATE-ID (Conditional Average Treatment Effects with Interference under Dynamic networks), a novel causal estimand that extends standard CATE to dynamic network settings where interference exists. They formally prove its identifiability under extended ignorability and consistency assumptions, addressing the theoretical challenge of recovering treatment effects from observational data in time-evolving networks.

10 retrieved papers
DSPNET framework for treatment effect estimation in dynamic networks

The authors develop DSPNET (Dynamic SPillover modeling NETwork), a novel framework that combines GCNs and RNNs to aggregate neighborhood and historical information for inferring dynamic hidden confounders. The framework learns dedicated interference representations to capture spillover effects and employs adversarial learning with gradient reversal to balance confounder representations across treatment groups.

10 retrieved papers
Interference representation learning via environment exposure

The authors introduce a method to model network interference by learning an interference representation that aggregates neighbors' treatments and hidden behavioral patterns, moving beyond simple scalar aggregations used in prior work. This representation serves as a data-driven embedding of the influence exerted by treated neighbors, capturing heterogeneous spillover effects in high-dimensional environments.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

CATE-ID estimand and identifiability for dynamic networks with interference

The authors introduce CATE-ID (Conditional Average Treatment Effects with Interference under Dynamic networks), a novel causal estimand that extends standard CATE to dynamic network settings where interference exists. They formally prove its identifiability under extended ignorability and consistency assumptions, addressing the theoretical challenge of recovering treatment effects from observational data in time-evolving networks.

Contribution

DSPNET framework for treatment effect estimation in dynamic networks

The authors develop DSPNET (Dynamic SPillover modeling NETwork), a novel framework that combines GCNs and RNNs to aggregate neighborhood and historical information for inferring dynamic hidden confounders. The framework learns dedicated interference representations to capture spillover effects and employs adversarial learning with gradient reversal to balance confounder representations across treatment groups.

Contribution

Interference representation learning via environment exposure

The authors introduce a method to model network interference by learning an interference representation that aggregates neighbors' treatments and hidden behavioral patterns, moving beyond simple scalar aggregations used in prior work. This representation serves as a data-driven embedding of the influence exerted by treated neighbors, capturing heterogeneous spillover effects in high-dimensional environments.