Hide/Show Apps

Improved link prediction for location based social networks with novel features and contextual feature reduction

Download
2019
Bayrak, Ahmet Engi
High penetration of broadband Internet access has made a revolution on the web usage, where users have become content generators rather than just consuming. People started to communicate, interact, maintain relationship and share data (image, video, note, location, etc.) with their acquaintances through varying online social network sites which are the key factors of that internet usage revolution. Online social networks with location sharing and interaction between people are called Location Based Social Networks (LBSNs). To use and benefit more from social networks, real life social links (friendship, acquaintanceship) should be represented well on them. Link Prediction problem has a motivation of studying social network evolution and trying to predict future possible links for representing the real-life relations better. In this work, we studied a comprehensive feature set which combines topological features with features calculated from temporal interaction data on LBSNs. We proposed novel features which are calculated by using time, category and common friend details of candidates and their social interaction in LBSNs. In addition, we proposed an effective feature reduction mechanism which helps to determine best feature subset in two steps. Contextual feature clustering is applied to remove redundant features and then a non-monotonic selection of relevant features from the calculated clusters are done by a custom designed genetic algorithm. Results depict that both new features and the proposed feature reduction method improved link prediction performance for LBSNs.