GraphFusion: Adaptive Label Enhancement with Heterophily-Aware Attention and Dynamic Residual Calibration

Rong Fu, Anqi Zhao, weizhi Tang, Simon James Fong

Keywords: graph neural networks, semi-supervised learning, label propagation, heterophily-aware attention, contrastive learning, dynamic residual calibration, graph transformers

TL;DR: A scalable graph learning framework combining contrastive pretraining, heterophily-aware attention, and dynamic label smoothing for robust node classification.

Abstract: We propose GraphFusion, a unified and scalable framework for graph-based semi-supervised learning that integrates high-order feature augmentation, contrastive self-supervised pretraining, and transformer-style attention. GraphFusion enhances node representations using multi-hop SIGN embeddings and contrastive alignment, followed by a lightweight Graphormer module with heterophily-aware attention. To improve label propagation, we introduce a dynamic residual calibration mechanism that adaptively adjusts smoothing strength per node based on model confidence. Our framework supports multi-label classification and ensemble fusion across multiple pathways. Extensive experiments on OGBN-ArXiv, OGBN-Products, and OGBN-Proteins demonstrate that GraphFusion outperforms strong baselines in both homophilous and heterophilous settings, while maintaining high efficiency and interpretability.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: true

Submission Guidelines: true

Anonymous Url: true

No Acknowledgement Section: true

Submission Number: 12902