International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Capsule Routing over Stacked GCN-GAT Embeddings with Negative Sampling for Graph Link Prediction

Authors :

Fatemeh Safari Sarvandi¹ Sayeh Mirzaei² Rooholah Abedian³

1- Faculty of Engineering Science, College of Engineering University of Tehran 2- Faculty of Engineering Science, College of Engineering University of Tehran 3- Faculty of Engineering Science, College of Engineering University of Tehran

Keywords :

Link Prediction،Graph Neural Networks،GCN،GAT،Capsule Routing،Negative Sampling

Abstract :

In this paper, we propose a novel multi-view graph-based link prediction framework that integrates Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and a Capsule Routing mechanism to capture rich and diverse node representations. While GCN and GAT provide complementary perspectives on neighborhood aggregation, their stacked embeddings are further refined through a capsule routing layer to preserve hierarchical spatial relationships and encode complex inter-node interactions. To enhance training and generalization, we adopt a negative sampling strategy to generate informative non-edge examples efficiently. We evaluate the proposed approach on five widely-used benchmark datasets: Cora, Citeseer, PubMed, Amazon Computers, and Amazon Photo. Experimental results demonstrate that our framework generally outperforms state-of-the-art GNN-based baselines, confirming the effectiveness of multi-view fusion and capsule-based refinement in link prediction tasks.