vowpal_wabbit

Why we included this project

Vowpal Wabbit is built for online learning, so each example updates the model the moment it arrives, which suits data that comes as a stream rather than a static training set. Its memory footprint stays bounded no matter how much data you feed it, thanks to the hashing trick, and the input format accepts free-form text as bag-of-words features split across namespaces. The built-in contextual bandit algorithms give it a home in interactive problems like ad ranking and personalized recommendation, and distributed allreduce training spreads the work across many machines. Teams already running production ranking pipelines may also appreciate the reduction framework, which makes it possible to experiment with learning-to-search and active-learning approaches without rebuilding your whole stack.