User Similarity Measures in Online Social Networks

Abstract

With the growing use of Online Social Networks (OSNs), these platforms have become central to engagement, participation, and community formation. As user bases expand, understanding user similarity becomes increasingly important for applications such as friend recommendation, community detection, and spam detection. Measuring user similarity includes determining behavioral patterns in addition to profile similarities and behavioral differences. In this research, we examine some classical similarity measures (Jaccard, Cosine, and Euclidean) versus deep learning-based measures, including Node Embeddings, Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and GraphSAGE with a synthetic Twitter dataset. User interactions are modeled as graphs and transformed into low-dimensional embeddings. Using the Deep Graph Infomax framework for unsupervised learning, clustering is applied and assessed through internal evaluation metrics: Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Score. The objective is to identify the most effective method for discovering meaningful user clusters in OSNs.