Combining Graph Representation Learning and Double Machine Learning for Adjusting Network Confounding in Static Networks

• In Event: Mathematical Sociology Section Flash Talks
  In Refereed Roundtable (60 minute): Table 2. Combining Graph Representation Learning and Double Machine Learning for Adjusting Network Confounding in Static Networks

Abstract

This study examines how network embeddedness is associated with confounding in regression models and proposes a method that combines graph representation learning (GRL) with double machine learning (DML) to adjust for such bias in static networks. Three sources of confounding are considered: homophily, contagion, and structural position. Node embeddings are learned using node2vec and incorporated as high-dimensional controls within a partially linear DML framework. Through simulations under homophily, homophily plus contagion, and mixed-membership structural position settings, the GRL+DML approach is compared with OLS, network autocorrelation models, and alternative specifications. Results show that embedding-based DML significantly reduces bias under homophily and structural position, and further improves performance when combined with autocorrelation controls in contagion settings. Although bias is not fully eliminated—particularly under structural position—the GRL+DML framework outperforms more conventional approaches, suggesting that learned node embeddings provide an effective proxy for latent network confounders.

Author

• Steve Liming Meng, Fudan University