
Research Article
attr2vec: Learning Node Representations from Attributes of Nodes
@INPROCEEDINGS{10.1007/978-3-030-69072-4_44, author={Pengkun Zheng and Yan Wen and Ming Chen and Geng Chen}, title={attr2vec: Learning Node Representations from Attributes of Nodes}, proceedings={Wireless and Satellite Systems. 11th EAI International Conference, WiSATS 2020, Nanjing, China, September 17-18, 2020, Proceedings, Part II}, proceedings_a={WISATS PART 2}, year={2021}, month={2}, keywords={Network embedding Unsupervised Feature learning}, doi={10.1007/978-3-030-69072-4_44} }
- Pengkun Zheng
Yan Wen
Ming Chen
Geng Chen
Year: 2021
attr2vec: Learning Node Representations from Attributes of Nodes
WISATS PART 2
Springer
DOI: 10.1007/978-3-030-69072-4_44
Abstract
In recent years, the research in the multiple fields of representation learning has led to the emergence of many excellent Network Embedding algorithms. Here we proposeattr2vec, a completely unsupervised algorithmic framework for learning the latent representations for nodes. Inattr2vec, we have adopted an attribute processing method similar to GCN, that is, taking the average of the attribute of the node’s neighbors as the attribute for the node. We also consider first-order neighbors and second-order neighbors separately to achieve an effect similar to multiple convolutional layers in GCN. In summary, our algorithm utilizes similar attribute processing idea of GCN, which can learn the graph topology and node attribute to generate latent representations for nodes, but implemented it with a completely unsupervised way. In some experiments on citation networks we demonstrate that our algorithm outperforms related unsupervised techniques by a significant margin.