Subgraph Generation for Generalizing on Out-of-Distribution Links
Jay Revolinsky, Harry Shomer, Jiliang Tang
TL;DR
FLEX is a novel graph generative model framework that improves out-of-distribution link prediction by structurally aligning sample distributions through adversarial co-training, without requiring expert knowledge.
Contribution
Introducing FLEX, a GGM framework that uses structural conditioning and adversarial co-training to enhance OOD link prediction without domain-specific tuning.
Findings
FLEX outperforms existing models in synthetic and real-world OOD settings.
Graph data augmentation improves link structure understanding.
FLEX does not require expert knowledge for different OOD scenarios.
Abstract
Graphs Neural Networks (GNNs) demonstrate high-performance on the link prediction (LP) task. However, these models often rely on all dataset samples being drawn from the same distribution. In addition, graph generative models (GGMs) show a pronounced ability to generate novel output graphs. Despite this, GGM applications remain largely limited to domain-specific tasks. To bridge this gap, we propose FLEX as a GGM framework which leverages two mechanism: (1) structurally-conditioned graph generation, and (2) adversarial co-training between an auto-encoder and GNN. As such, FLEX ensures structural-alignment between sample distributions to enhance link-prediction performance in out-of-distribution (OOD) scenarios. Notably, FLEX does not require expert knowledge to function in different OOD scenarios. Numerous experiments are conducted in synthetic and real-world OOD settings to demonstrate…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The motivation is clear and intuitive. 2. The evaluation is conducted on four datasets with diverse shift schemes, and an ablation study is also provided.
1. I’m not fully convinced by the performance improvements shown in Tables 1 and 3. AUC scores below 0.5–0.6 are essentially trivial, indicating near-random predictions. Although FLEX often yields statistically significant gains, improvements such as 50% → 52% provide limited practical utility. The per-dataset breakdown in Table 3 further shows that GCN+FLEX frequently increases AUC while remaining in the trivial range. In some cases (e.g., Backward–PA), FLEX even degrades GCN’s AUC from 73% to
1. OOD link prediction is a critical bottleneck for GNN deployment, yet existing work predominantly focuses on node/graph classification. This paper explicitly demonstrates that standard OOD generalization methods (e.g., IRM, CORAL) exhibit limited effectiveness in LP tasks (from Table 1), providing empirical evidence and establishing a robust problem motivation. 2. The paper employs k-hop subgraph generation instead of full-graph generation, introduces a labeling trick to enable GNNs to perce
1. The paper repeatedly cites Pearl's causal framework, yet the FLEX generation process does not model structural equation models or interventions. Instead, it achieves structural differences solely through KL divergence maximization. This approach aligns more closely with diversity sampling in data augmentation than with counterfactuals in a strict causal sense. 2. Gamma significantly impacts performance, but selection relies on grid search. In real-world out-of-distribution scenarios where the
By reformulating the task in terms of structural feature distributions (e.g., Common Neighbors), the paper provides a principled explanation for why traditional link predictors underperform under distribution shifts. The proposed set-theoretic and ELBO-based analysis forms the unified theoretical perspectives on OOD generalization for link prediction. The empirical evaluation spans multiple benchmark datasets and diverse graph structures, consistently demonstrating the robustness and OOD genera
Although the appendix provides a set-theoretic argument that counterfactual subgraph generation can enlarge the overlap between training and testing distributions, the analysis remains qualitative. It lacks a quantitative derivation of generalization bounds, risk functions, or error guarantees. The theoretical foundation that KL divergence regularization and structural diversity objectives necessarily improve OOD generalization remains insufficient. The paper emphasizes generating “structurally
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Advanced Clustering Algorithms Research · Semantic Web and Ontologies