Extending 'rvw' and reintegrating vowpal wabbit

Background

Vowpal Wabbit (or vw for short) is a machine learning system which pushes the frontier of machine learning with techniques such as online, hashing, allreduce, reductions, learning2search, active, and interactive learning. Basic and initial R integration exist, more should be added.

Related work

The RVowpalWabbit packgage embeds an older copy of vw. It is on CRAN and available everywhere, but no longer current.

The rvw package (on GitHub) uses an external vw binary. This provides a current baseline.

Details of your coding project

Possible goals

• Extend rvw and complete the data import function (vw uses a particular format, we need one or more converters)
• More ambitious: Rethink and rework RVowpalWabbit now that an external library libvw exists. Provide an updated package built against current library sources.
• For either/both: Continue best practice of benchmarking against existing vw code and results.

Expected impact

Machine Learning is a very important (and, dare we say, hot) topic. Vowpal Wabbit is an established high-performance backend with many interesting features, and the focus of ongoing research. A tighter coupling to R would be very welcome.

Mentors

Students, please contact mentors below after completing at least one of the tests below.

• Dirk Eddelbuettel [email protected] is involved with the two R packages mentioned above, and then some including Rcpp.
• James Balamuta [email protected] is graduate student and expert in R, Rcpp, and lots of other things.

Tests

Students, please do one or more of the following tests before contacting the mentors above.

• Easy: Add one or more additional examples or benchmarks.
• Medium: Start towards a new data.importer for rvw.
• Hard: Start towards a new R and libvw integration using Rcpp. Build a basic package that does at least one thing and benchmark / compare to reference output from one of the existing solutions.

Solutions of tests

Students, please post a link to your test results here.

Ivan Pavlov (https://github.com/ivan-pavlov/rvw-tests2018)