Ultra-fast and Efficient Network Embedding for Gigascale Biological Datasets

Graph/network representation learning (or graph/network embedding) is a widely used machine learning technique in industry recommending systems and has recently been applied in computational biology. Popular network representation learning algorithms include random walk and matrix factorization methods, but they do not scale well to large networks. To accommodate the fast growth of real-world network datasets, especially biological datasets, we engineered and improved several network embedding algorithms via intensive computational optimization (e.g., randomized-generalized singular value decomposition/SVD, efficient sketching via ProbMinHash including edge weights) and parallelization to allow ultra-fast and accurate embedding of large- scale networks. We present GraphEmbed, a computer program for scalable, memory-efficient network embedding. GraphEmbed can perform embedding for large-scale networks with several billion nodes in less than 2 hours on a commodity computing cluster. We benchmark it against standard datasets and demonstrate consistent speed and accuracy advantages over state-of-the- art techniques. We also propose centric AUC, a new metric for evaluating link-prediction accuracy in network embedding. It corrects the bias in conventional AUC caused by the highly skewed node degree distributions, which are typically found in real-world networks, especially biological networks. Taken together, GraphEmbed solves a major challenge in large-scale network representation learning for networks in general and biological networks in particular.