Abstract: Network representation learning is an important research topic, which purpose is to represent high-dimensional attribute networks as low-dimensional dense vectors, so as to provide effective feature representation for the next task. SNE, a recently proposed attribute network representation learning model, uses both network structure and attribute information to represent learning network nodes. However, this model belongs to the unsupervised model and cannot make full use of some easily acquired prior information to improve the quality of the learned feature representation. Based on the above considerations, a semi-supervised attribute network representation learning method SSNE is proposed, which takes the attribute network and a small number of node priors as input of the feedforward neural network, learning the optimal node representation by maintaining the network link structure and a small number of node priors in the output layer through multiple hidden layer nonlinear transformations. Compared with the existing mainstream methods on four real attribute networks and two artificial attribute networks, the results show that the representation learned in this method has better performance in clustering and classification tasks.

Key words: attribute networks, semi-supervised learning, representation learning, clustering

摘要： 网络表示学习是一个重要的研究课题，其目的是将高维的属性网络表示为低维稠密的向量，为下一步任务提供有效特征表示。最近提出的属性网络表示学习模型SNE（Social Network Embedding）同时使用网络结构与属性信息学习网络节点表示，但该模型属于无监督模型，不能充分利用一些容易获取的先验信息来提高所学特征表示的质量。基于上述考虑提出了一种半监督属性网络表示学习方法SSNE（Semi-supervised Social Network Embedding），该方法以属性网络和少量节点先验作为前馈神经网络输入，经过多个隐层非线性变换，在输出层通过保持网络链接结构和少量节点先验，学习最优化的节点表示。在四个真实属性网络和两个人工属性网络上，同现有主流方法进行对比，结果表明本方法学到的表示，在聚类和分类任务上具有较好的性能。

关键词: 属性网络, 半监督学习, 表示学习, 聚类