Learning Multigraph Node Embeddings Using Guided Lévy Flights

Learning efficient representation of graphs has recently been studied extensively for simple networks to facilitate various downstream applications. In this paper, we deal with a more generalized graph structure, called multigraph (multiple edges of different types connecting a pair of nodes) and propose Multigraph2Vec, a random walk based framework for learning multigraph network representation. Multigraph2Vec samples a heterogeneous neighborhood structure for each node by preserving the inter-layer interactions. It employs Lévy flight random walk strategy, which allows the random walker to travel across multiple layers and reach far-off nodes in a single step. The transition probabilities are learned in a supervised fashion as a function of node attributes (metadata based and/or network structure based). We compare Multigraph2Vec with four state-of-the-art baselines after suitably adopting to our setting on four datasets. Multigraph2Vec outperforms others in the task of link prediction, by beating the best baseline with 5.977% higher AUC score; while in the multi-class node classification task, it beats the best baseline with 5.28% higher accuracy. We also deployed Multigraph2Vec for friend recommendation on Hike Messenger.