A novel hybrid approach for link prediction problem in social network

Abstract

Link prediction in dynamic social network is an inherent challenging task to find out the intrinsic associations among the objects. This association can be discovered with the help of pattern of connection in already linked objects. This paper addresses the three novel hybrid feature extraction methodologies that leverage the significant insight of existing model available for link prediction. The proposed technique offers better agreement of causal link to be happened between pair of objects in future. The conventional link prediction models are applying features which are typically grounded on the node cardinality. These predictor models are common neighbour predictor, Adamic/Adar, Jaccard's coefficient, preferential attachment, friends measures. So, hybrid feature extraction techniques based on principle of feature ensemble is introduced to generate predictive model to be learned for link prediction. Jacccard's coefficient and preferential attachment (JCPA) model, Jaccard's coefficient and Adamic/Adar (JCAA) model and preferential attachment and Adamic/Adar (PAAA) model are generated with the context of ascertaining more accuracy in the performance of link prediction classifier. The probabilistic model based on Naive Bayes classifier is used as a base classier. Experimental results shows that we have achieved more accuracy in classifier performance which also been verified using receiver operating characteristic (ROC) curve.