Semantic Similarity Measures for Topological Link Prediction

The semantic approach to data linked in social networks uses information extracted from node attributes to quantify the similarity between nodes. In contrast, the topological approach exploits the structural information of the network, e.g., nodes degree, paths, neighbourhood breadth. For a long time, such approaches have been considered substantially separated. In recent years, following the widespread of social media, an increasing focus has been dedicated to understanding how complex networks develop, following the human phenomena they represent, considering both the meaning of the node and the links structure and distribution. The link prediction problem, aiming at predicting how networks evolve in terms of connections between entities, is suitable to apply semantic similarity measures to a topological domain. In this paper, we introduce a novel topological formulation of semantic measures, e.g., NGD, PMI, Confidence, in a unifying framework for link prediction in social graphs, providing results of systematic experiments. We validate the approach discussing the prediction capability on widely accepted data sets, comparing the performance of the topological formulation of semantic measures to the conventional metrics generally used in literature.