Estimating treatment effects from observational data is a fundamental challenge in many domains. Most existing methods rely on the unconfoundedness assumption, which is often violated in practice due to latent confounders. To address this issue, recent studies incorporate auxiliary network information to infer such variables. However, they typically treat all observed features and network structure as proxies for confounders, without distinguishing among different latent factors. This entanglement may lead to biased estimates when latent factors that affect treatment and outcome in different ways are not properly distinguished. To overcome this limitation, we propose TNDGVA, a variational autoencoder framework for estimating individual treatment effects from networked observational data. TNDGVA leverages network information to learn disentangled latent representations that are explicitly decomposed into instrumental, confounding, adjustment, and noise factors, and incorporates an independence regularization based on the Hilbert–Schmidt Independence Criterion to promote effective disentanglement. The learned representations are used to model potential outcomes and improve causal effect estimation. Experiments on multiple networked datasets demonstrate the effectiveness of TNDGVA for treatment effect estimation.
Fan et al. (Thu,) studied this question.