Nodeseq2vec: Node Sequence Representation Learning

Current network representation learning techniques focus on mapping each network node to a low-dimensional space, essentially functioning as node representation learning approaches. In this paper, we introduce a novel method for network representation, called node sequence representation learning, which shows highly competitive performance compared to traditional node representation approaches. We develop two scalable frameworks, nodeseq2vec and nodeseq2vec++ aim to project each node sequence generated from a network to a low-dimensional space inspired by distributed representations of documents. Therefore, the node se17quences which may represent a research field, a cooperative group, etc., on network are computable. The nodeseq2vec framework is able to learn high-quality node sequence feature representations through modeling the relationship between a node sequence and all the nodes it has. The nodeseq2vec++ framework further combines semantics with structure in net- work with external text information. In our experiments, we design two methods to approximately represent node features through multiple node sequences, and we comprehensively evaluate our framework against existing network representation models on various real-world datasets across different domains. Ultimately, we achieve significantly improved performance compared to existing state-of-the-art methods in both multi-label classification and citation recommendation tasks.