Abstract:

In heterogeneous social networks, the prediction of co-authorship is a representative prediction of relationship, which is quite different from the link prediction method of homogeneous networks in node representation, network construction and so on. Fully considering the unique information of meta-path and node attribute characteristics of heterogeneous networks, a composite vectorization representation of nodes is proposed, which combines the TF-IDF features of nodes with network embedding based on Metapath2vec. In the representation of meta-path, the same type of nodes in meta-path are merged and reconstructed to further extract the implicit information between the same type of nodes in meta-path. Also, the prediction of co-authorship in bibliographic network is realized by Convolutional Neural Network（CNN）. Experimental results show that the method of complex vectorization of nodes and reconstruction of meta-path can better characterize heterogeneous social networks, as well as, better evaluation indexes of prediction are given compared with other methods.

Key words: link prediction, co-authorship, heterogeneous social network, semantic attributes, meta-path

摘要：

在异构社会网络中，合著关系的预测是具有代表性的一类关系预测，与同构网络的链接预测方法在节点表示、网络构造等方面存在较大差异。综合考虑异构社会网络特有的元路径信息和节点属性特征，提出了节点的复合向量化表示：将节点的TF-IDF特征与基于Metapath2vec算法的向量化表示相结合；在元路径的表示上采取元路径中同类型节点归并重构的方法，以提取元路径中同类型节点间的隐含信息；并通过卷积神经网络（CNN）实现学术网络的合著关系预测。实验结果表明，节点的复合向量化表示及重构元路径方法可以更好地表征异构社会网络，与其他方法对比中均获得更好的预测评价指标。

关键词: 链接预测, 合著关系, 异构社会网络, 语义属性, 元路径