Abstract: In recent years, social networks have become increasingly hot, and data mining based on social networks has also arisen. Link prediction (LP) is an important topic of network data mining, which uses the known network structure and other information to predict and estimate the possibility of linking between two nodes that are not yet linked. Link prediction in social network can be used to recommend friends, filter redundant information, improve user’s satisfaction and loyalty, and build a healthy social networking environment. In previous researches, attentions are focused on structure information or node attributes, in order to analyze the global or local properties. Considering the natures of microblog social network, this paper proposes a link prediction method combining multiple features which includes node features, topological features, social features and voting features. Based on these features, 4 machine learning algorithms, SVM, naive Bayes, random forest and logical regression, are applied on microblog social network data to train predictive models to predict potential social links. The results show that combining multiple features performs better than the traditional features, and the combination of multiple features can achieve highest accuracy.

Key words: link prediction, data mining, features extraction, friends recommendation, machine learning algorithm