Distributed Representations of Subgraphs

There has been a surge in research interest in learning feature representation of networks in recent times. Researchers, motivated by the recent successes of embeddings in natural language processing and advances in deep learning, have explored various means for network embedding. Network embedding is useful as it can exploit off-the-shelf machine learning algorithms for network mining tasks like node classification and link prediction. However, most recent works focus on learning feature representation of nodes, which are ill-suited to tasks such as community detection which are intuitively dependent on subgraphs. In this work, we formulate a novel subgraph embedding problem based on an intuitive property of subgraphs and propose SubVec, an unsupervised scalable algorithm to learn feature representations of arbitrary subgraphs. We demonstrate usability of features learned by SubVec by leveraging them for community detection problem, where it significantly out performs non-trivial baselines. We also conduct case-studies in two distinct domains to demonstrate wide applicability of SubVec.