Link Prediction in Social Networks: Combining Topological and Contextual Data in a Community Detection Based Method

Link Prediction (LP) is the task of predicting which nodes in a network will interact in the future. A common approach to LP is to compute degrees of compatibility between unconnected node pairs in the network. In such an approach, the predictive model uses some similarity metrics applied in the same way for all pairs of unconnected nodes, independent of the positions those nodes have in the network structure. More recent work has applied a different approach: they first detect communities in the network and then apply LP to each community. Nevertheless, these works have an important limitation: their community detection process only considers topological aspects of the network. They fail to consider, at the time of node grouping, characteristics related to the application context, such as participant's profiles, interests, and preferences, which may be fundamental both for the identification of more cohesive communities and for a greater assertiveness in predicting new connections. This paper proposes a method for LP that uses a community detection phase that combines topological and contextual data. This community detection phase takes into account characteristics of the network's nodes in order to separate them into groups whose internal content is cohesive. Tests with twelve scenarios of four networks popularly used in LP studies provided experimental evidence that the proposed method can overcome the state-of-the-art contextual data agnostic community detection based LP methods.