Abstract: Link prediction is the key issue in social network analysis, which aims to predict new links from the known networks. In real networks, there exist a large number of unlinked node pairs, whose intrinsic information can be helpful to the link prediction task. In this paper, link prediction is looked as a binary classification problem, where the semi-supervised learning techniques are used to take advantage of unlabeled data in learning process. Two semi-supervised paradigms are used：self-training and co-training. The experimental results in real datasets Enron and DBLP show that, exploiting unlabeled data can effectively improve the accuracy of the link prediction.

Key words: link prediction, semi-supervised learning, self-training, co-training, social network analysis

摘要： 链接预测是社会网络分析领域的关键问题，研究如何从已知网络中预测可能存在的新链接。现实网络中存在了大量未连接的节点对，从中挖掘潜在信息可以帮助实现链接预测任务。将链接预测视为二类分类问题，使用半监督学习技术，利用网络中的未标记数据帮助学习。使用了两种半监督范式：自我训练和协同训练。在现实数据集Enron和DBLP中的实验结果表明，链接预测任务中采用未标记数据能够有效提高预测的准确率。

关键词: 链接预测, 半监督学习, 自我训练, 协同训练, 社会网络分析