[Paper Review] Graph Neural Networks for Graphs with Heterophily: A Survey

Xin Zheng, Wang, Yi+6 more|arXiv (Cornell University)|Feb 14, 2022

Advanced Graph Neural NetworksComputer Science87 citations

TL;DR

This survey provides a comprehensive taxonomy and benchmarks for GNNs on heterophilic graphs, detailing non-local neighbor extension and architecture refinement approaches, plus future directions.

ABSTRACT

Recent years have witnessed fast developments of graph neural networks (GNNs) that have benefited myriads of graph analytic tasks and applications. In general, most GNNs depend on the homophily assumption that nodes belonging to the same class are more likely to be connected. However, as a ubiquitous graph property in numerous real-world scenarios, heterophily, i.e., nodes with different labels tend to be linked, significantly limits the performance of tailor-made homophilic GNNs. Hence, GNNs for heterophilic graphs are gaining increasing research attention to enhance graph learning with heterophily. In this paper, we provide a comprehensive review of GNNs for heterophilic graphs. Specifically, we propose a systematic taxonomy that essentially governs existing heterophilic GNN models, along with a general summary and detailed analysis. Furthermore, we discuss the correlation between graph heterophily and various graph research domains, aiming to facilitate the development of more effective GNNs across a spectrum of practical applications and learning tasks in the graph research community. In the end, we point out the potential directions to advance and stimulate more future research and applications on heterophilic graph learning with GNNs.

Motivation & Objective

• Motivate the study of GNNs on heterophilic graphs where homophily assumptions fail.
• Provide a systematic taxonomy categorizing heterophilic GNN approaches into non-local neighbor extension and GNN architecture refinement.
• Summarize real-world benchmarks and datasets for heterophilic graphs to support fair evaluation.
• Analyze limitations and propose future directions in interpretability, robustness, scalability, and data exploration.

Proposed method

• Propose a systematic taxonomy of heterophilic GNNs based on how neighbors are defined and how messages are aggregated.
• Review two main categories: non-local neighbor extension (high-order neighbor mixing and potential neighbor discovery) and GNN architecture refinement (adaptive aggregation, ego-neighbor separation, inter-layer combination).
• Summarize representative models and their techniques, including spectral and spatial aggregation schemes, and different ways to handle heterophily in messages.
• Compile and summarize real-world heterophilic benchmarks and their statistics to support robust evaluation.

Experimental results

Research questions

• RQ1 What are the current method families for GNNs on heterophilic graphs and how do they address neighbor discovery and information aggregation?
• RQ2 How do non-local neighbor extension and GNN architectural refinements compare and complement each other in heterophilic settings?
• RQ3 What benchmarks exist for heterophilic GNN evaluation and what limitations do they have?
• RQ4 What directions are most promising for improving interpretability, robustness, scalability, and data exploration in heterophilic GNNs?

Key findings

• A systematic taxonomy divides heterophilic GNNs into non-local neighbor extension and GNN architecture refinement.
• Non-local methods include high-order neighbor mixing and potential neighbor discovery to capture informative distant nodes.
• Architecture refinement methods cover adaptive aggregation, ego-neighbor separation, and inter-layer combination to enhance discriminability.
• A set of real-world heterophilic benchmarks (e.g., WebKB subsets, Chameleon, Squirrel, Wiki, ArXiv-Year, Snap-Patents, etc.) are summarized to support evaluation.
• The survey highlights interpretability, robustness, scalability, and comprehensive benchmarks as key future directions for heterophilic GNN research.