Recommending friends in local social networks: An envelope of algorithms

As the proliferating of online social networks, a lot of researchers studied the topics of link prediction, and developed a plenty of friend recommending algorithms. The researchers try to illustrate that their algorithms perform better than other algorithms, especially in terms of prediction accuracy. However, it is much possible that no friend recommending algorithm can really beat the other algorithms based on the same data/information. Specially, in this study, we find that the friend recommending algorithms perform differently when we evaluate the algorithms using different performance indicators, or when the algorithms are applied for different groups of users classified by the user degree. Based on these findings, we propose a method to construct envelops of algorithms, to combine the "best part" of each algorithm together for specified purpose. We collect data from a Chinese dominant social network website to compare the existing algorithms and to verify the proposed envelope method. The results show that the proposed algorithm can enhance the performance of friend recommending algorithms for specified purposes.