Graph Neural Networks for Graphs With Heterophily: A Survey
Xin Zheng, Yi Wang, Yixin Liu, Ming Li, Miao Zhang, Di Jin et al. · 2026 · 109 citationsRead the paper
Recent years have witnessed fast developments of graph neural networks (GNNs) that have benefited myriad graph analytic tasks and applications. Most GNNs rely 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 graphsare 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 governs existing heterophilic GNN models, along with general summaries and detailed analyses. Furthermore, we discuss the relationship between 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 potential directions to advance and inspire future research and applications on heterophilic graph learning with GNNs.
What the full text shows
Working prototype- Heterophily-specific GNN designs generally outperform homophilic GNNs on heterophilic graphs, though well-tuned GCNs can succeed if neighborhood distributions across classes remain distinguishable. (Heterophilic graphs)
- Ego-neighbor separation (detaching self-loops and using non-mixing operations like concatenation) improves distinguishable node embedding learning. (Heterophilic graphs)
- Prior-based methods (e.g., CPGNN) capture both heterophily and homophily by propagating prior belief estimations based on a compatibility matrix. (Class-information based weight estimation)
- Attention-based mechanisms (e.g., CGCN) allow for signed messages (positive for same class, negative for opposing classes) to handle dissimilar neighbors. (Cosine similarity and relative node degrees)
- Setup
- Comprehensive review of existing literature and models (e.g., H2GCN, CPGNN, NLGNN, RAW-GNN) across multiple years (2018-2025).
- How
- The paper provides a systematic taxonomy and survey of GNNs for heterophilic graphs, categorizing them into prior-based, attention-based, hybrid weight, and neighbor extension methods (local-level and global-level incorporation). It analyzes various architectural refinements such as ego-neighbor separation, higher-order neighborhood mixing, and dynamic routing.
- Limitations
- Early empirical results in the field may be flawed due to class imbalance and data leakage. · Traditional metrics hinder cross-dataset comparisons. · Homophilic GNNs fail when neighborhood distributions across distinct classes become indistinguishable.
Summarized by Gemma 4 from the open-access full text · 12 Oct 2026. Read it