Recommendation API Improvements

[Michael Ekstrand]

I have now merged a set of Top-N recommendation API changes. After fleshing out documentation & testing, I think we will be ready to release 0.6, which should stabilize the rest of the APIs.

The new change introduces a 'CandidateSelector' algorithm type, and the standard TopN recommender implementation uses a Predictor and a CandidateSelector. The CandidateSelector needs to be fit (this is how it memorizes the user's rated items). In a production setting, you could substitute a candidate selector that accesses a live database.

If you provide a set of candidates to the 'recommend' method, it will override the candidate selector.

The batch recommend interface is unchanged from 0.5, except that 'candidates' is now optional. It will still 'adapt' a Predictor, but if you do that and *don't* provide 'candidates', you'll have a problem since it can't fit the candidate selector. LensKit will emit a warning in this case. MultiEval adapts predictors before training, so you can provide predictors to it and they will be trained and used as recommenders properly.

The GettingStarted doc is updated: http://lkpy.lenskit.org/en/latest/GettingStarted.html

--

Michael D. Ekstrand — michael...@boisestate.edu — https://md.ekstrandom.net
Assistant Professor, Dept. of Computer Science, Boise State University

People and Information Research Team (PIReT) — http://coen.boisestate.edu/piret/