AFIF: Automatically Finding Important Features in community evolution prediction for dynamic social networks

Abstract

Applications of Intelligent systems have increased the impact of modeling and analyzing in dynamic social networks. For effective decision making, models must be able to forecast the outcome of each option and determine which option is the best for a particular situation. In this context, community evolution prediction is a challenging and time-consuming task, which is the extraction of structural features from large real-world networks. We present AFIF, Automatically Finding Important Features, an efficient solution to examine communities’ structural features and also to find a proper subset of promising features in order to predict the upcoming changes of social networks. AFIF combines two key concepts to find prominent features: (i) Prioritization of attributes based on their Spearman’s correlation with other features. This enables us to know the features that can represent the rest and to explore which features are unique compared to others. (ii) Training a boosting learner and prioritizing attributes based on their usage frequency in learning process to realize which features are more valuable. Eventually, important features are determined by random forest classifier. We have then conducted extensive experiments and confirmed that our selection of features delivers an outstanding performance in contrast with using the entire set of features.