Abstract
We describe an efficient, scalable machine learning library that enables veryfast training of generalized linear models. We demonstrate that our library canremove the training time as a bottleneck for machine learning workloads,opening the door to a range of new applications. For instance, it allows moreagile development, faster and more fine-grained exploration of thehyper-parameter space, enables scaling to massive datasets and makes frequentre-training of models possible in order to adapt to events as they occur. Ourlibrary, named Snap Machine Learning (Snap ML), combines recent advances inmachine learning systems and algorithms in a nested manner to reflect thehierarchical architecture of modern distributed systems. This allows us toeffectively leverage available network, memory and heterogeneous computeresources. On a terabyte-scale publicly available dataset forclick-through-rate prediction in computational advertising, we demonstrate thetraining of a logistic regression classifier in 1.53 minutes, a 46x improvementover the fastest reported performance.