Community detection in Multimedia Social Networks using an attributed graph model

In this paper, we design a novel data model for a Multimedia Social Network, that has been modeled as an attribute graph for integrating semantic analysis of multimedia content published by users. It combines features inferred from object detection, image classification, and hashtag analysis in a unified model to characterize a user from different points of view. On top of this model, community detection algorithms have been applied to unveil users’ communities. Hence, we design a framework integrating multimedia features with different community detection approaches (topological, deep learning, representation learning, and game theory-based) to improve detection effectiveness. The proposed framework has been evaluated on a real-world dataset, composed of 4.5 million profiles publishing more than 42 million posts and 1.2 million images, to investigate the impact of different features on both graph-building and community detection tasks. The main findings of the proposed analysis show how combining different sets of features inferred from multimedia content allows to achieve the highest modularity score w.r.t. other configurations although it requires a higher running time for building the underlined network. Specifically, representation and game theory-based algorithms achieve the highest results in terms of Modularity measure by exploiting the semantic and contextual information integrated into the proposed model.