SINE: Side Information Network Embedding.

Network embedding learns low-dimensional features for nodes in a network, which benefits the downstream tasks like link prediction and node classification. Real-world networks are often accompanied with rich side information, such as attributes and labels, while most of the efforts on network embedding are devoted to preserving the pure network structure. Integrating side information is a challenging task since the effects of different attributes vary with nodes and the unlabeled nodes can be influenced by diverse labels from neighbors, not to mention the heterogeneity and incompleteness. To overcome this issue, we propose Side Information Network Embedding (SINE), a novel and flexible framework using multiple side information to learn a node representation. SINE defines a flexible and semantical neighborhood to model the inscape of each node and designs a random walk scheme to explore this neighborhood. It can incorporate different attributes information with particular emphasis depending on the characteristics of each node. And label information can be both explicitly and potentially integrated into the representation. We evaluate our method and existing state-of-the-art methods on the tasks of multi-class classification. The experimental results on 5 real-world datasets demonstrate that our method outperforms other methods on the networks with side information.