A Novel RNN-based Ensemble Model for Link Prediction on Dynamic Social Network

Book excerpt: Social networking sites have gained much popularity in the recent years. With millions of people connected virtually generate loads of data to be analyzed to infer meaningful associations among links, arising in different applications ranging from recommendation systems to social networks. Most of the existing methods predict interactions between individuals for static networks, ignoring the dynamic features of social networks. In this paper, we describe the most popular similarity indices and compare their performance in their ability to show links with the highest probability of being added from the initial network. Moreover, we propose a novel RNN-based ensemble model for link prediction on a dynamic social network. We ensemble some similarity indices to obtain a higher performance of link prediction. It applies long short-term memory (LSTM) neural network to predict the possibility of potential links of social network. Finally, we evaluate the link prediction performances of our proposed method and 11 similarity indices with different accuracy measures. ex: AUC, precision, recall, error rate. We also discuss window size influence with AUC. Experimental results show that our method can always find an ensemble with better accuracy than all similarity indices regardless of the dataset.