Causally Fair Node Classification on Non-IID Graph Data
Yucong Dai, Lu Zhang, Yaowei Hu, Susan Gauch, Yongkai Wu
TL;DR
This paper introduces a causality-based approach for fair node classification on non-IID graph data, addressing biases by estimating interventional distributions with a novel message passing variational autoencoder.
Contribution
It proposes the MPVA model that leverages causal inference and $do$-calculus to improve fairness in non-IID graph data, a setting often overlooked in prior work.
Findings
MPVA outperforms traditional methods in bias mitigation
Effective approximation of interventional distributions demonstrated
Addresses fairness in complex, interconnected data environments
Abstract
Fair machine learning seeks to identify and mitigate biases in predictions against unfavorable populations characterized by demographic attributes, such as race and gender. Recently, a few works have extended fairness to graph data, such as social networks, but most of them neglect the causal relationships among data instances. This paper addresses the prevalent challenge in fairness-aware ML algorithms, which typically assume Independent and Identically Distributed (IID) data. We tackle the overlooked domain of non-IID, graph-based settings where data instances are interconnected, influencing the outcomes of fairness interventions. We base our research on the Network Structural Causal Model (NSCM) framework and posit two main assumptions: Decomposability and Graph Independence, which enable the computation of interventional distributions in non-IID settings using the -calculus.…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. Identifies two general conditions (decomposability, graph independence) under which *do‑calculus* extends to non‑IID graphs, yielding a *closed‑form interventional expression* (Theorem 3). This is a nice result IMO. 2. Uses WL colors to ground the structural summarization that enables the graph isomorphism logic with causal identification. 3. Semi‑synthetic experiments include ground‑truth interventions and show close agreement. Reasonable sensitivity analysis and assumption‑violation stud
1. **Graph Independence** (exogenous $U_X$ independent of structure C) is unlikely in many social graphs with homophily or formation dynamics. The paper shows degradation under violations (Table 4) but does not provide diagnostics or robustness strategies on real data. Does it make sense to add stress tests on real graphs (e.g. degree- or community‑stratified analyses)? 2. **Scope of interference patterns.** Most empirical work assumes interference flows only from S to X (Fig. 1e), whereas rea
1. The paper tackles the underexplored problem of causal fairness in non-IID graph settings. 2. By focusing on interventional distributions rather than counterfactuals, the method sidesteps well-known identifiability problems in counterfactual inference.
1. While the paper establishes theoretical conditions for when $do$-calculus applies, the experiments do not systematically verify that these conditions hold for the datasets used. 2. The paper assumes a specific causal structure where the sensitive attribute $S$ affects node features $X$ through an intermediate variable $A$, which then affects the outcome $Y$ (i.e., $S \rightarrow A \rightarrow X \rightarrow Y$). However, this structure fails to account for direct causal relationships that comm
1. Novelty: The paper tackles a novel and underexplored problem: defining a causal fairness notion for non-IID (graph-structured) data. This direction is timely and relevant, as traditional fairness definitions based on the i.i.d. assumption fail in networked settings. 2. Significance: The proposed framework has strong potential impact, since non-IID fairness is essential for a variety of graph-based learning applications, such as social network analysis and credit risk prediction. 3. Quality: T
1. Some notations and statements could be better explained. The model figure is not very informative, and a clearer diagram would help improve readability. 2. The paper would benefit from additional illustration or intuitive explanation of the theoretical assumptions (e.g. Decomposability and Graph Independence) to make their implications more accessible. 3. It remains unclear whether the proposed MPVA framework scales to large graphs or how sensitive its performance is to the accuracy of the es
- The paper’s main theoretical contribution is indeed the extension of causal inference from the IID assumption to graph-structured, networked data via the network structural causal model. - The integration of message passing neural networks with a variational autoencoder provides a coherent deep learning implementation for estimating interventional distributions. - Results show that the proposed MPVA performs better than the baseline fair GNNs included.
- The decomposability and graph independence conditions are unlikely to hold in most real-world graphs. Real graphs often exhibit dependencies between the node structure and exogenous variables (e.g., social homophily), limiting the scope of the theoretical framework. - Although the model claims to estimate interventional distributions, it does not analyze what these causal effects reveal about bias sources in the data - It does not thoroughly discuss computational complexity or the scalabilit
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Privacy-Preserving Technologies in Data
MethodsBalanced Selection · Causal inference