GEMSEC: Graph Embedding with Self Clustering.
Network embedding procedures localize nodes of a graph in a low dimensional feature space, which enables machine learning on graph data. In this paper we propose GEMSEC - a graph embedding algorithm which learns a clustering of the nodes simultaneously with the embedding. Clusters or communities are natural constructs in social networks, and incorporating this objective is found to produce more representative embeddings than traditional methods. We find that GEMSEC extracts high quality clusters on real world social networks and outperforms existing embedding algorithms on clustering tasks. The embedding procedure preserves the neighborhoods of individual vertices, while at the same time, it forms k different clusters for a given k. GEMSEC generalizes earlier works in the domain. It augments the skip-gram like embedding procedures and is agnostic of the neighborhood sampling strategy. In experiments, the joint embedding and clustering is found to be extremely scalable, with computation costs linear in the size of the input.
Publisher URL: http://arxiv.org/abs/1802.03997