OpenNovelty.org

H3^3GNNs: Harmonizing Heterophily and Homophily in GNNs via Self-Supervised Node Encoding

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Graph Neural NetworksHeterophily and HomophilySelf Supervise Learning
Abstract:

Graph Neural Networks (GNNs) have made significant advances in representation learning on various types of graph-structured data. However, GNNs struggle to simultaneously model heterophily and homophily, a challenge that is amplified under self-supervised learning (SSL) where no labels are available to guide the training process. This paper presents H3^3GNNs, an end-to-end graph SSL framework designed to harmonize heterophily and homophily through two complementary innovative perspectives: (i) Representation Harmonization via Joint Structural Node Encoding. Nodes are embedded into a unified latent space that retains both node specificity and graph structural awareness for harmonizing heterophily and homophily. Node specificity is learned via linear and non-linear node feature projections. Graph structural awareness is learned via a proposed Weighted Graph Convolutional Network (WGCN). A self-attention module enables the model learning-to-adapt to varying levels of patterns. (ii) Objective Harmonization via Predictive Architecture with Node-Difficulty–Aware Masking. A teacher network processes the full graph. A student network receives a partially masked graph. The student is trained end-to-end, while the teacher is an exponential moving average of the student. The proxy task is to train the student to predict the teacher’s embeddings for all nodes (masked and unmasked). To keep the objective informative across the graph, two masking strategies that guide selection toward currently hard nodes while retaining exploration are proposed. Theoretical underpinnings of H3^3GNNs are also analyzed in detail. Comprehensive evaluations on benchmarks demonstrate that H3^3GNNs achieves state-of-the-art performance on heterophilic graphs (e.g., +7.1% on Texas, +9.6% on Roman-Empire over the prior art) while matching SOTA on homophilic graphs, and delivering strong computational efficiency.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes H³GNNs, a self-supervised learning framework that harmonizes heterophily and homophily through joint structural node encoding and a predictive architecture with node-difficulty-aware masking. It resides in the 'Joint Structural Encoding and Self-Supervised Learning' leaf, which contains only three papers total (including this one). This is a relatively sparse research direction within the broader taxonomy of fifty papers, suggesting the specific combination of structural encoding with self-supervised objectives for mixed homophily-heterophily graphs remains an emerging area rather than a crowded subfield.

The taxonomy reveals several neighboring approaches. The sibling leaf 'Dual-View and Multi-View Contrastive Frameworks' contains three papers focusing on complementary views via contrastive objectives, while 'Frequency-Based Signal Decomposition' (two papers) uses spectral techniques to separate graph signals. Adjacent branches include 'Decoupled Representation Learning' (two papers) that separately model homophilic and heterophilic structures, and 'Homophily-Aware Augmentation and Edge Manipulation' (three papers) that design augmentation strategies based on connection patterns. H³GNNs diverges from these by integrating structural awareness directly into node embeddings rather than relying on view construction, spectral decomposition, or augmentation strategies.

Among twenty-one candidates examined, four refutable pairs were identified across three contributions. The 'Representation Harmonization' contribution examined ten candidates with one appearing to provide overlapping prior work, while nine remain non-refutable or unclear. The 'Objective Harmonization' contribution examined only one candidate, which was refutable. The 'Weighted Graph Convolutional Network' examined ten candidates, with two refutable and eight non-refutable. This limited search scope suggests that while some prior work exists in weighted convolutions and predictive architectures, the joint structural encoding approach may retain novelty within the examined literature.

Based on top-twenty-one semantic matches and citation expansion, the analysis covers a focused subset of the field rather than exhaustive prior work. The sparse population of the target taxonomy leaf and the moderate refutation rate across contributions suggest the work occupies a relatively underexplored intersection of structural encoding and self-supervised learning for mixed homophily-heterophily graphs, though definitive novelty claims require broader literature coverage beyond this limited search.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
21
Contribution Candidate Papers Compared
4
Refutable Paper

Research Landscape Overview

Core task: self-supervised learning on graphs with mixed homophily and heterophily. Real-world graphs often exhibit both homophilic regions (where connected nodes share similar labels) and heterophilic regions (where neighbors differ), posing challenges for traditional graph neural networks that assume uniform connectivity patterns. The taxonomy organizes the field into several main branches: Homophily-Heterophily Adaptation Mechanisms develop methods that explicitly adjust to varying local connectivity patterns, such as adaptive filtering or decoupled architectures like Decoupled Non-Homophilous[6]; Heterogeneous Graph Self-Supervised Learning tackles graphs with multiple node and edge types; Self-Supervised Learning for Heterophilic Graphs focuses on contrastive and generative techniques tailored to heterophilic settings, exemplified by Heterophily Self-Supervised[2] and Beyond Homophily[3]; Universal and Augmentation-Free Self-Supervised Frameworks aim for broadly applicable methods that avoid hand-crafted augmentations; Homophily-Centric Self-Supervised Methods leverage strong homophily assumptions when appropriate, as in HomoGCL[29]; and Domain-Specific, Graph-Level, and Specialized Learning Paradigms address particular application contexts or auxiliary tasks. A particularly active line of work explores how to jointly encode structural properties and learn representations without relying on a single homophily assumption. For instance, Heterogeneous Homophily View[1] and Heterophilic Recurring Pattern[4] investigate recurring motifs and local homophily variations, while SimGCL[5] and related contrastive approaches experiment with noise-based augmentations that remain robust across diverse connectivity regimes. H3GNNs[0] sits within the Joint Structural Encoding and Self-Supervised Learning cluster, closely related to HGNNs Joint Encoding[31] and H3GNNs Joint Encoding[49], which all emphasize integrating explicit structural encodings with self-supervised objectives to handle mixed homophily-heterophily scenarios. Compared to purely contrastive methods like SimGCL[5] or decoupled architectures like Decoupled Non-Homophilous[6], H3GNNs[0] focuses on enriching node representations with structural context before applying self-supervised learning, aiming to capture both local and global graph properties in a unified framework.

Claimed Contributions

Representation Harmonization via Joint Structural Node Encoding

The authors propose a joint structural node encoding scheme that combines linear and non-linear node feature projections to preserve node specificity with a Weighted Graph Convolutional Network (WGCN) to capture graph structural awareness. A self-attention module enables adaptive learning across varying levels of homophily and heterophily patterns.

10 retrieved papers
Can Refute
Objective Harmonization via Predictive Architecture with Node-Difficulty–Aware Masking

The authors introduce a teacher-student predictive framework where the teacher provides stable holistic node representations from the full graph while the student learns from a partially masked graph. Two dynamic masking strategies guide selection toward currently hard nodes, ensuring the learning objective harmonizes easy and hard samples as well as homophilic and heterophilic signals.

1 retrieved paper
Can Refute
Weighted Graph Convolutional Network (WGCN)

The authors propose WGCN, which learns edge weights dynamically to adaptively control message passing. This balances smoothing and sharpening operations, enabling the model to handle diverse graph structures more effectively while preserving high efficiency and avoiding oversmoothing in heterophilic regions.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Representation Harmonization via Joint Structural Node Encoding

The authors propose a joint structural node encoding scheme that combines linear and non-linear node feature projections to preserve node specificity with a Weighted Graph Convolutional Network (WGCN) to capture graph structural awareness. A self-attention module enables adaptive learning across varying levels of homophily and heterophily patterns.

Contribution

Objective Harmonization via Predictive Architecture with Node-Difficulty–Aware Masking

The authors introduce a teacher-student predictive framework where the teacher provides stable holistic node representations from the full graph while the student learns from a partially masked graph. Two dynamic masking strategies guide selection toward currently hard nodes, ensuring the learning objective harmonizes easy and hard samples as well as homophilic and heterophilic signals.

Contribution

Weighted Graph Convolutional Network (WGCN)

The authors propose WGCN, which learns edge weights dynamically to adaptively control message passing. This balances smoothing and sharpening operations, enabling the model to handle diverse graph structures more effectively while preserving high efficiency and avoiding oversmoothing in heterophilic regions.

Table of Contents