Time-aware methods for Link Prediction in Social Networks

Online social network services has become one of the most influential and key source of service providing, information/knowledge sharing and many other Internet based activities. The rapid growth of social networks shows the increasing popularity of these services among the users. The growth of social networks occurs as a result of adding new users and new links between users. The emergence of new links has primacy in the study of social network evolution. Thus, predicting/recommending future links in social networks has attracted a great deal of attention. Link prediction has many applications and, it offers many benefits to the users of social networking services such as providing fast and accurate recommendations or suggestions to the users. However, highly structured massive real-world networks involving heterogeneous entities with complex associations have added new challenges to link prediction research due to different factors such as sparsity, complexity, size, time-dependent nature of the networks. There have been numerous attempts to address the problem of link prediction through diverse approaches. Most common way is to measure the closeness/similarity of nodes to each other in terms of different social aspects. These similarities change over time due to highly dynamic behavior of social networks. The existing static similarity measures have not been able to cope with rapidly evolving social networks thus, are not sufficient for accurate link prediction. In order to alleviate this problem, we contributed by introducing two novel timeaware features, 1) Time score which is capable of dealing with temporality of common neighbors and, 2) T_Flow computes information flow between nodes by considering link activeness which vary over time. We used the latest timestamps of interactions/links to compute them. The novel features used in conjunction with supervised machine learning method for link prediction. Both methods tested on real world social networks namely, facebook friendship network data and coauthorship data extracted from ePrint archives. The results revealed a significant improvement in link prediction accuracy for both features comparing with the existing features.