Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Distribution and Label Consistency for Graph Out-of-Distribution Generalization

Song Wang, Xiaodong Yang, Rashidul Islam, Huiyuan Chen, Minghua Xu, Jundong Li, Yiwei Cai

TL;DR

This work tackles graph out-of-distribution generalization under distribution shifts by addressing two core inconsistencies: distribution consistency in augmented graphs and label consistency in invariant subgraphs. It introduces DLG, a unified framework that learns edge-masked modifiers to generate both augmented graphs $G_a$ and invariant graphs $G_v$ from a given graph, guided by distribution- and label-consistency losses. Through mutual-information-based augmentation and a label-guided invariance objective, DLG achieves superior performance on diverse graph-level and node-level OOD benchmarks, outperforming established baselines. The approach demonstrates that jointly optimizing augmentation and invariance with informative regularization yields robust generalization while preserving supervision signals, suggesting practical benefits for real-world graph learning under shifts.

Abstract

To deal with distribution shifts in graph data, various graph out-of-distribution (OOD) generalization techniques have been recently proposed. These methods often employ a two-step strategy that first creates augmented environments and subsequently identifies invariant subgraphs to improve generalizability. Nevertheless, this approach could be suboptimal from the perspective of consistency. First, the process of augmenting environments by altering the graphs while preserving labels may lead to graphs that are not realistic or meaningfully related to the origin distribution, thus lacking distribution consistency. Second, the extracted subgraphs are obtained from directly modifying graphs, and may not necessarily maintain a consistent predictive relationship with their labels, thereby impacting label consistency. In response to these challenges, we introduce an innovative approach that aims to enhance these two types of consistency for graph OOD generalization. We propose a modifier to obtain both augmented and invariant graphs in a unified manner. With the augmented graphs, we enrich the training data without compromising the integrity of label-graph relationships. The label consistency enhancement in our framework further preserves the supervision information in the invariant graph. We conduct extensive experiments on real-world datasets to demonstrate the superiority of our framework over other state-of-the-art baselines.

Enhancing Distribution and Label Consistency for Graph Out-of-Distribution Generalization

TL;DR

This work tackles graph out-of-distribution generalization under distribution shifts by addressing two core inconsistencies: distribution consistency in augmented graphs and label consistency in invariant subgraphs. It introduces DLG, a unified framework that learns edge-masked modifiers to generate both augmented graphs and invariant graphs from a given graph, guided by distribution- and label-consistency losses. Through mutual-information-based augmentation and a label-guided invariance objective, DLG achieves superior performance on diverse graph-level and node-level OOD benchmarks, outperforming established baselines. The approach demonstrates that jointly optimizing augmentation and invariance with informative regularization yields robust generalization while preserving supervision signals, suggesting practical benefits for real-world graph learning under shifts.

Abstract

To deal with distribution shifts in graph data, various graph out-of-distribution (OOD) generalization techniques have been recently proposed. These methods often employ a two-step strategy that first creates augmented environments and subsequently identifies invariant subgraphs to improve generalizability. Nevertheless, this approach could be suboptimal from the perspective of consistency. First, the process of augmenting environments by altering the graphs while preserving labels may lead to graphs that are not realistic or meaningfully related to the origin distribution, thus lacking distribution consistency. Second, the extracted subgraphs are obtained from directly modifying graphs, and may not necessarily maintain a consistent predictive relationship with their labels, thereby impacting label consistency. In response to these challenges, we introduce an innovative approach that aims to enhance these two types of consistency for graph OOD generalization. We propose a modifier to obtain both augmented and invariant graphs in a unified manner. With the augmented graphs, we enrich the training data without compromising the integrity of label-graph relationships. The label consistency enhancement in our framework further preserves the supervision information in the invariant graph. We conduct extensive experiments on real-world datasets to demonstrate the superiority of our framework over other state-of-the-art baselines.
Paper Structure (14 sections, 12 equations, 2 figures, 1 table)

This paper contains 14 sections, 12 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Problem Formulation
  4. Methodology
  5. Modification
  6. Distribution Consistency Enhancement
  7. Label Consistency Enhancement
  8. Optimization
  9. Experiments
  10. Datasets and Settings
  11. Results of Graph-Level OOD Generalization
  12. Ablation Study of DLG
  13. Effect of Diversity Loss and Support Set Size
  14. Conclusion

Figures (2)

  • Figure 1: The overall framework of DLG. Given a graph $G=(\mathbf{A}, \mathbf{X})$, we feed it into $\text{GNN}_v$ to learn a graph representation, which will be used to generate an invariant graph $G_v$ for classification. Meanwhile, the input graph is also processed by $\text{GNN}_a$ to generate an augmented graph $G_a$, which shares the same label as $G_v$ but with different structures and edges. To enhance label consistency, we propose the classification loss $\mathcal{L}_c$ for optimization. To enhance distribution consistency, we design a loss $\mathcal{L}_d$ that leverages randomly sampled graphs from existing data to ensure $G_a$ aligns with distributions of existing data.
  • Figure 2: The performance of our framework DLG with different modules removed.