采用主动学习的动态网络链接预测方法

Using active learning in link prediction for dynamic networks

针对网络动态性和稀疏性的特点,在网络进化及链接预测过程中引入主动学习范式,提出了一种新的动态网络链接预测方法。首先为网络中每个结构特征的变化序列都生成一个分类器,再用这些分类器对每个未连接的节点对进行评分并把预测结果差异较大的节点对样本交于用户判别;一旦获取真实的标记(即节点间是否存在链接),系统采用更新的训练集重新训练各分类器并整合得到最终的模型。在三个现实的合著者网络数据集中的实验表明,在动态网络链接预测方法中引入主动学习在AUC值指标上有显著提高。

In view of the characteristics of network dynamics and sparsity,introducing the active learning paradigm in the process of network evolution and link prediction,this paper proposed a new dynamic network link prediction method.The method firstly generated a classifier for the variation sequence of each structural feature.Then it used these classifiers to score each unconnected node pair and gave the node with the large difference of prediction result to the user.Once the real labels were obtained(i.e.,whether the links existed),these classifiers would be retrained using the updated training set and be integrated to the final model.Experimental results in three real co-author network datasets show that the performance of link prediction method for dynamic networks can significantly improve the AUC measure by using active learning.