We propose a graph embedding method, Conv2Vec, that is based on graph convolutions and trained with unsupervised learning with objectives derived from concepts in classical graph algorithms. Conv2Vec is a versatile embedding method as it can embed both plain graphs and graphs with node features, and it can utilize many variants of graph convolutions.Comparing with the random walk plus SkipGram approach for embedding plain graphs, Conv2Vec achieves accuracies similar to Node2Vec with Cora and Citeseer, and outperforms Node2Vec for Pubmed. Interestingly, direct learning on the graph with both GCN and GraphSAGE performs poorly. This suggests that convolution alone does not effectively capture global features in the graph and motivates our search for alternative objectives for learning global structural features.
Citation and DOI:
Versatile feature learning with graph convolutions and graph structures, 3rd IEEE ICDM Workshop on DEEP LEARNING AND CLUSTERING, In conjunction with IEEE ICDM 2021 December 7-10, 2021. Auckland, New Zealand
