Features fusion based link prediction in dynamic neworks

Abstract Link prediction is the process of detecting new or missing links in any network. This work focuses on link prediction in dynamic networks. One of the major challenges of the link prediction problem is to achieve high accuracy. One of the well-known methods for link prediction is the similarity-based method, which uses similarity-based score. The three widely used similarity-based indices are Local (L), Global (G), and Quasi-local (Q) for calculating similarity scores. In this work, we have proposed a model, namely the LGQ model. In our LGQ model, the wide categories of indices are used in different combinations(L, G, Q, LG, LQ, GQ, LGQ) for feature set generation that can be used with various machine learning techniques for link prediction. In local similarity indices, we have considered Common Neighbors (CN), Adamic/Adar Index (AA), Jaccard Coefficient (JC), and Preferential Attachment (PA). In global similarity indices, we have considered cos+, Average Commute Time (ACT), Shortest Path (SP), MFI (Matrix Forest Index), and in quasi-local indices, local path Index (LP), Path of Length 3 (L3). We have tested and validated our proposed model by conducting numerical simulations on six well-known dynamic network datasets. The results demonstrate that the proposed LGQ model and its variations outperform several baseline algorithms in terms of AUPR, F1 score, Balanced accuracy score (BAC), and AUC. We also demonstrate and contrast the preference of Neural Network and Xgboost based prediction models for different variations of proposed feature sets.