MG2Vec+: A multi-headed graph attention network for multigraph embedding

Aman Roy, Shravika Mittal, Tanmoy Chakraborty

Abstract: Representation learning of graphs in the form of graph embeddings is an extensively studies area, especially for simple networks, to help with different downstream applications such as node clustering, link prediction, and node classification. In this paper, we propose MG2Vec+, a method that generates node embeddings for a multigraph, a network structure comprising multiple types of edges between pairs of nodes. MG2Vec+ uses multi-headed attention layers to aggregate multiple types of edge-relations that can exist among nodes. The parameters are learned using a graph likelihood loss function which ensures that the sum of attention scores for high-priority nodes is larger as compared to low-priority nodes. We compare MG2Vec+ with nine existing baseline methods after modifying them to our setting on four real-world datasets. MG2Vec+ outperforms the competing methods when evaluated on two downstream tasks: (1) link prediction, and (2) multi-class node classification. It is able to achieve a 5.88% higher AUC-ROC score than the best baseline for link prediction and 9.52% higher classification accuracy than the best baseline for the multi-class node classification task. The superiority of MG2Vec+ can be explained by its principled way of capturing multi-relational contexts and learning them in an unsupervised manner with the same set of parameters using graph likelihood loss.