Graph Representation Learning

Abstract

Graph-based representation learning is pivotal in deriving low-dimensional vectors from nodes and edges, enhancing performance in numerous tasks. This study primarily focuses on evaluating the efficacy of various graph-based GNN models for real-world applications, such as clustering and classification. By leveraging contrastive learning, we obtain robust representations of graphs using datasets like Cora, Citeseer, and PubMed. Our methodology emphasises the integration of graph neural networks with advanced graph augmentation techniques, setting it apart from traditional graph embedding methods, including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and GraphSAGE. The crux of our approach is embedding data with both structural and semantic similarities in close proximity, enhancing the quality and clarity of representations for downstream applications. We employ a cosine similarity metric to evaluate graph similarity, thus gauging the efficacy of the augmentation strategies. Executed using the reliable PyTorch Geometric library, our research aspires to drive advancements in graph representation learning by juxtaposing various GNN models via contrastive learning for clustering and classification tasks. Moreover, we spotlight salient factors influencing the proficiency of these techniques, providing valuable insights for refining real-world applications leveraging graph-based machine learning.