Your go-to online machine learning library
Vowpal Wabbit is a fast and flexible learning system that empowers you to test and explore real data and find meaningful outcomes to real problems.
Expand the boundaries of
online machine learning
Vowpal Wabbit is sponsored by Microsoft Research and influenced by a growing ecosystem of community contributions, academic research, and proven algorithms. It provides fast, efficient, and flexible online machine learning techniques for reinforcement learning, supervised learning, and more.
Optimize rewards with Contextual Bandits
Contextual Bandits are a form of reinforcement learning(RL). The vast majority of production real-world RL systems prefer contextual bandits techniques.
For an overview of a contextual bandit problem and how to use Vowpal Wabbit in a CB setting with Python.
Learning to Search(L2S) builds a policy for sequential decision problems that optimizes global value with training time advice. L2S is a form of guided reinforcement learning, which enjoys global guarantees on performance. Typical applications are in natural language processing but the technique is much more widely applicable.
In active learning, a learner chooses which examples to label given what it knows and unlabeled examples to choose from. Active learning is used to minimize the cost of labelling datasets.
Extreme multi-class or multi-label learning addresses learning a classifier that can tag each data point with the most relevant label or subset of labels, from an extremely large label set. It has applications in document tagging, as well as ranking and recommendation systems by reformulating them as multi-label learning tasks, where each item to be ranked or recommended is considered as a separate label.
Most learning algorithms require all data before they start learning. Vowpal Wabbit(VW) additionally enables efficient learning from a data source which continuously grows. This allows VW to be used in situations where a problem changes over time or in situations where interactive learning is required.
VW handles learning problems with any number of sparse features. Vowpal Wabbit was the first published tera-scale learner.
VW achieves scalability using distributed out-of-core learning. Together with the hashing techniques that VW pioneered, this means the memory footprint of VW is bounded independent of the training set.
The input format for VW is substantially more flexible than many other toolkits allow. Examples can have features consisting of free form text and multiple feature sources can be explicitly represented within an example and later used.
VW includes example manipulators for efficiency and ease of deployment. Different feature sources can be suppressed or combined to maximize performance.
When solving real world problems the most important step is framing the problem effectively. VW enables this by supporting solutions to a wide range of problems through reductions to common learning algorithms. This enables you to focus on framing the problem to achieve the best solution.
Install Vowpal Wabbit
A vehicle for machine learning research
Vowpal Wabbit’s design and capabilities are informed and influenced by advanced research and proven algorithms. Researchers who wish to implement their algorithm in Vowpal Wabbit have access to quick deployment and extensive use.
Join our community
Browse the code and file issues
Get answers to your questions
Chat with the team and community