Directed Homophily-Aware GNN Shows Up to 15% Gains on Directed Graph Tasks

Global: Directed Homophily-Aware GNN Shows Up to 15% Gains on Directed Graph Tasks

Researchers introduced the Directed Homophily-aware Graph Neural Network (DHGNN) in a paper posted on arXiv, aiming to improve performance on directed graphs that exhibit heterophilic neighborhoods. The framework incorporates homophily‑aware and direction‑sensitive components, employs a resettable gating mechanism, and integrates node representations from both original and reverse directions. In benchmark experiments, DHGNN surpassed state‑of‑the‑art methods, achieving up to a 15.07% improvement in link prediction accuracy.

Background on Graph Neural Networks

Graph Neural Networks have become a cornerstone for learning from graph‑structured data, excelling in tasks such as node classification and link prediction. However, many existing models assume homophily—where connected nodes share similar attributes—and often overlook the directional nature of edges, limiting their effectiveness on heterophilic and asymmetric graphs.

Innovations Introduced by DHGNN

DHGNN addresses these gaps through two key innovations. First, a resettable gating mechanism dynamically adjusts the influence of incoming messages based on measured homophily levels and the informativeness of each neighbor. Second, a structure‑aware, noise‑tolerant fusion module combines representations derived from forward and reverse traversals of the graph, preserving directional information.

Experimental Evaluation

The authors evaluated DHGNN on a suite of directed graph datasets that span both homophilic and heterophilic regimes. Experiments covered two primary tasks: node classification, which predicts node labels, and link prediction, which forecasts the existence of edges. Comparative baselines included leading GNN architectures that do not explicitly model directionality or heterophily.

Performance Outcomes

Across all tested datasets, DHGNN consistently outperformed competing methods. Notably, the model achieved a 15.07% relative gain over the strongest baseline in link prediction, while also delivering measurable improvements in node classification accuracy. These results suggest that accounting for directional homophily can substantially enhance predictive capabilities.

Insights from Model Analysis

Further analysis revealed that the gating mechanism effectively captures directional homophily gaps and adapts to fluctuating homophily across network layers. This behavior provides a clearer understanding of how messages propagate in complex graph structures and highlights the importance of adaptive message weighting.

Potential Applications and Future Directions

By delivering robust performance on directed and heterophilic graphs, DHGNN opens avenues for applications in domains such as citation networks, social media interaction graphs, and biological pathways where edge direction carries critical meaning. Future research may explore scaling the framework to larger graphs and integrating additional contextual features.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.