Causal and Structural Inference for Heterogeneous Network Interference

About This Grant

Network interference is a fundamental driving force behind phenomena such as intervention spillover, behavioral contagion, and information diffusion in interconnected systems. In this project, the investigator aims to develop statistical methods to understand and quantify whether, and to what extent, an individual’s behavior or status is influenced by others through network interactions. This project focuses on two central challenges in estimating network interference from observational data: individual heterogeneity and network confounding. To overcome the challenges, the investigator will develop an expressive and interpretable statistical model that adaptively learn how the effect of interference depends on individual-level contextual and network information. This model is expected to provide insights and an analytics tool towards the mechanisms of network interference. In addition, the investigator will design novel estimators to measure the causal effects of network interference, leveraging advanced machine learning techniques to address complex confounding among connected individuals and to make efficient use of limited experimental data. The methodologies developed in this project will advance the fields of causal inference and graph-based machine learning. The tools developed will have broad applicability to network data in social science, public health, political science, economics, and business, and will support new theoretical developments in areas such as social influence, disease transmission, marketing, cultural evolution, and collective behavior. The project will provide research opportunities for graduate students. In this project, the investigator aims to (1) design a data-driven framework for estimating heterogeneous spillover effects using graph neural networks; (2) develop intervention effect estimation method that integrates active learning to address sample size limitation which is common in real applications; and (3) construct a directed graphical model to identify latent propagation patterns in heterogeneous network cascades. The methodological foundation consists of two main innovations: an attention-based neural network model for robustly estimating individual exposure mapping, and a network mixture model for recovering diffusion structures at the population level from cascade data. To further improve estimation efficiency, the project introduces novel data augmentation strategies that leverage contextual information and network structure, enhancing the causal estimation accuracy of the intervention effect, even in data-scarce settings. The main advantage of the new methods is the decomposition of target estimands into two components: a global network-based interference structure and local individual heterogeneity. The latter is approximated using advanced graph machine learning techniques, enabling the model to strike a balance between expressiveness and interpretability. Overall, this research will provide theoretical and computational tools for studying network interference, leading to the development of open-source software tailored for practical applications across various disciplines. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.