Evolution Similarity for Dynamic Link Prediction in Longitudinal Networks

Abstract

Link prediction problem in network science has spawned not only over myriad applications but also experienced extensive methodological improvements. Different link prediction methods perform feature engineering to build different topological or nodal attribute based metrics measuring the similarity/proximity between non-connected actor pairs to deal with the inference of future associations among them. On the contrary, dynamic link prediction methods have catered the evolutionary process and network dynamics of longitudinal networks. Evolution similarity between node pairs (e.g., similarity between rates of acquiring neighbours by actor pairs over time) can be considered to generate dynamic metrics for the purpose of dynamic link prediction in longitudinal networks. In this study, we attempt to build dynamic similarity metrics by considering the similarity between temporal evolutions of non-connected actor pairs. For this purpose, this study utilises time series forecasting methods to model the temporal evolution of actors’ network positions/importance and then it utilizes a dynamic programming method to determine the similarity between these evolutions of actor pairs to quantify the likelihood of future associations among them. Supervised link prediction models exploiting these dynamic similarity metrics were built and performances were compared against some baseline static metrics (i.e., common neighbours). High performance scores achieved by these features, examined in this study, represent them as prospective candidates not only for dynamic link prediction task but also in various applications like security and recommender systems.