Link prediction based on spatio-temporal networks

Abstract

In a network composed of complex entities and relationships, its topology offers multiple patterns that could indicate different properties such as importance, rank, and category. These properties indicate new trends through their deeper analysis. A selection of social activities and interactions are not only dynamic, but their strength and reach evolve over time and locality. Anticipating the likelihood of future social interactions is similar to the Link Prediction Problem. This paper describes how a social network was built from a snapshot of a spatiotemporal dataset which includes user identifier, geo-coded location, and time event as attributes for people checking-in certain localities. This social network is then used as a basis to predict the likelihood of two persons checking-in at the same place over a comparable time interval. A set of features is used to hold scores indicating the similarities of pairs of nodes. One of the prediction features employed in this study is a time-based weight which describes the activeness of the network nodes based on how recent their adjacent interactions are. A supervised binary classification technique is used with these features on part of the dataset to segment results based on whether a link is formulated in the future. The model created is then used on a distinct test set to generate predictions. The results of this empirical study, yielded an overall accuracy of more than 90%. Other performance measures such as precision, positive prediction rate, and negative prediction rate are also used to aid the model’s evaluation