Graph Disentangle Causal Model: Enhancing Causal Inference in Networked Observational Data

TL;DR

The paper tackles estimating individualized treatment effects in networked observational data where hidden confounders exist. It introduces the Graph Disentangle Causal Model (GDC), a three-part framework with a causal disentangle module, a three-path graph aggregation module, and a causal constraint module to learn true adjustment and confounder factors and their counterfactuals. By separating features into adjustment and confounding components and aggregating them with treatment-aware graph attention, GDC yields more accurate factual and counterfactual predictions, validated on semi-synthetic BlogCatalog and Flickr datasets where it outperforms baselines. This approach offers a principled way to mitigate confounding bias in graphs while preserving predictive power, with implications for personalized treatment and recommendations in networked domains.

Abstract

Estimating individual treatment effects (ITE) from observational data is a critical task across various domains. However, many existing works on ITE estimation overlook the influence of hidden confounders, which remain unobserved at the individual unit level. To address this limitation, researchers have utilized graph neural networks to aggregate neighbors' features to capture the hidden confounders and mitigate confounding bias by minimizing the discrepancy of confounder representations between the treated and control groups. Despite the success of these approaches, practical scenarios often treat all features as confounders and involve substantial differences in feature distributions between the treated and control groups. Confusing the adjustment and confounder and enforcing strict balance on the confounder representations could potentially undermine the effectiveness of outcome prediction. To mitigate this issue, we propose a novel framework called the \textit{Graph Disentangle Causal model} (GDC) to conduct ITE estimation in the network setting. GDC utilizes a causal disentangle module to separate unit features into adjustment and confounder representations. Then we design a graph aggregation module consisting of three distinct graph aggregators to obtain adjustment, confounder, and counterfactual confounder representations. Finally, a causal constraint module is employed to enforce the disentangled representations as true causal factors. The effectiveness of our proposed method is demonstrated by conducting comprehensive experiments on two networked datasets.

Figures (5)

• Figure 1: An illustration comparing conventional ITE estimation methods on networked data with our proposed approach. Upper right: Conventional methods treat all features as confounder without identifying causal factors. Strict distribution balance may undermine outcome prediction efficacy. Lower right: Upon disentangling causal factors, we find that adjustment distributions between two groups can be effectively balanced, while confounder distributions cannot. Instead of enforcing distribution balance, we introduce a counterfactual mapping process to enhance ITE estimation.
• Figure 2: Causal graph for networked observational data. $\mathbf{X}_i, \mathbf{X}_j, \mathbf{X}_k$ represent the features of unit $i$, neighboring unit $j$ with the same treatment as unit $i$, and neighboring unit $k$ with a different treatment, respectively. $\mathbf{X}_{a, \cdot}$, $\mathbf{X}_{c, \cdot}$ are their corresponding disentangled adjustment and confounder from their own features. And $\mathbf{E}_{a, i}, \mathbf{E}_{c, i}, \mathbf{E}_{cf, i}$ are the adjustment, confounder, and counterfactual confounder by incorporating network information. $\mathbf{Y}_{f,i}, \mathbf{Y}_{cf,i}$ are factual outcome and counterfactual outcome.
• Figure 3: Overall architecture of our proposed Graph Disentangle Causal Model.
• Figure 4: Experimental results of parameter studies. The $\sqrt{\epsilon_{PEHE}}$ initially decreases and then increases as the values of $\mathcal{W}_1$,$\mathcal{W}_2$,$\mathcal{W}_3$ increase.
• Figure 5: T-SNE projections of learned confounder and adjustment embeddings.