Release notes for LensKit 3.0

The Git changelog and the list of closed tickets and pull requests provide more information on what has happened, including bugs that have been fixed.

LensKit 3.0 is the upcoming major release of LensKit. This release will not be backwards-compatible with LensKit 2.2. These release notes are not exhaustive, but are intended to provide a general guide to the changes to look for. As 3.0’s final release nears, we will be writing additional documentation on translating and updating experiments and other migration concerns.

For a preview of the manual for LensKit 3.0, see here.

Organization Changes

• LensKit code has been moved from org.grouplens.lenskit to org.lenskit.
• The LensKit Maven group has moved from org.grouplens.lenskit to org.lenskit.
• The LensKit Data Structures module is being removed in favor of fewer custom data structures. The custom data structures still in use are in the LensKit Core.
• We have standardized our Builder and Provider language. Components that build other components in the recommender build process are called Providers, and the term Builder is reserved for classes that are used to programmatically assemble other objects.

New Public APIs

The LensKit public APIs (ItemScorer, RatingPredictor, etc.) have been redesigned with new data structures for reporting results.

New Data Access Layer

LensKit’s Data Access Layer has been completely rewritten to make it easier to link to many different of data sources. Documentation is forthcoming.

Algorithm Changes

• FunkSVD no longer supports predict-time updating. This logic was not well-tested and always disabled in recommended configurations; this release completely removes it. We hope to re-introduce it in a better-designed and more configurable manner later.

• If FunkSVD cannot find an item vector for an item, it now declines to score that item rather than returning the baseline score. Baseline scores can be added with a fallback item scorer.

• The quantized rating predictor has moved to lenskit-predict.

• The rating predictor no longer takes a baseline scorer to use as a fallback when the item scorer cannot produce scores, to reduce duplication of logic with the fallback item scorer. The default behavior is still the same, however, as it uses a @PredictionScorer qualifier whose default implementation is FallbackItemScorer.

Evaluator Changes

‘Evaluation scripts’ have gone away in favor of discrete command line tools (train-test, crossfold) and a Gradle plugin for scripting them. This means we can leverage all of Gradle’s engineering in developing a DSL for task automation.

Metric Changes

• Incompatible Change: The Precision and Recall metrics now count users for whom no recommendations could be generated as having precision and recall of 0 in the global average.
• The MRR.OfGood variant has gone away, there is now just MRR.