Structural Entanglement in Dynamic Networks

• In Event: New Advances in Social Network Analysis

Abstract

Distinguishing social influence (e.g., contagion) from social selection (e.g., homophily) remains the central identification challenge in network sociology. Traditional approaches face a stark trade-off: Stochastic Actor-Oriented Models (SAOMs) offer rigorous identification but fail to scale, while modern Graph Neural Networks (GNNs) scale effortlessly but suffer from "structural entanglement," conflating attribute-driven sorting with position-driven constraint. To solve the scalable and interpretable issue, I introduce a Dynamic Disentangled Network DML (DD-GNN-DML) framework, operationalizing a hybrid identification strategy: (1) within-unit first-differencing to eliminate time-invariant unobserved heterogeneity (e.g., latent personality), and (2) orthogonal bounding via degree-preserving randomization to identify bounds on the remaining time-varying confounding. Simulation results ($N=10,000$) demonstrate that while standard GNN baselines remain biased by latent homophily (Bias $>70\%$), my framework recovers valid bounds for the causal peer effect, bridging the gap between rigorous sociological identification and scalable deep learning.

Author

• Yuanhao (Colin) Liu, Johns Hopkins University