Link Prediction in Dynamic Real-Life Friendship Networks: A Case Study

Link prediction in social networks plays an important role in mining and analyzing the evolution of social networks. However, most of the previous works are based on online social networks, which usually represent human's social behaviors on the artificially designed platform. In this paper, we examine the problem of evaluating the accuracy of different link prediction methods in the dynamic real-life friendship networks, which represents 172 first-year undergraduate students' friendship link formation process during 4 months. We evaluate the performance of 10 neighbor-based methods, 2 path-based methods and 2 random-walk-based methods. We explored how prediction performances change with the chosen methods, networks' growing and underlying mechanism in the real-life friendship networks. We find most link prediction methods can achieve high accuracies in a very difficult situation, where a great part of links vanished, and an enormous amount of links were created. There are significant differences in the performance according to the type of link prediction method. The accuracies increase with the friendship networks growing. KATZ and CST overcome other methods with long-term predictive capability, high robustness, and well performance. At last, most methods both work well on the reciprocal network and non-reciprocal network.