It is is an open-source fast out-of-core learning system originally developed by John Langford at Yahoo! Research, and currently being developed at Microsoft Research. VW is one of the offered ML options in Microsoft Azure. It is notable for its many features including e.g., reduction functions, importance weighting, selection of different loss functions, optimization algorithms. VW has been used to learn a tera-feature (1012) data-set on 1000 nodes in one hour, and can run properly in single machine, Hadoop and HPC cluster.

Strong points

• Open source, efficient, scalable and fast out-of-core online learning supported by strong IT companies (Microsoft, previously Yahoo).
• Feature identities are converted to a weight index via a hash using 32-bit MurmurHash3 (the hashing trick).
• Exploiting multi-core CPUs on Hadoop cluster by own MPI-AllReduce library, parsing of input and learning are done in separate threads.
• Allows using non-linear features e.g. n-grams.
• Product of the strong industrial laboratory, compiled C++ code, well-maintained (github), well-supported.

Weak points

• The number of available ML methods is sufficient but limited.
• API-oriented environment only.

Return to Contemt