Release 0.11

The primary focus of this release is making LensKit use Grapht rather than PicoContainer, with additional refactorings of components to more effectively take advantage of Grapht. We also made significant improvements to how the evaluator works to allow evaluation scripts to be more flexible, and made a number of other improvements and bug fixes.

There are a number of backwards-incompatible changes in this release. Configurations and algorithm implementations will definitely need to be updated; other client code may need updates as well.

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

General Changes

  • Made SparseVector return VectorEntry objects when iterated, rather than Long2DoubleMap.Entry instances. All code that iterates over sparse vectors will need to be updated.
  • Added methods to ItemScorer and GlobalItemScorer to output scores to sparse vectors.
  • Begun documenting API stability (#103).
  • Events are no longer Cloneable, and the clone method is now a deprecated alias for copy().

Algorithm Changes

  • The default user similarity for user-user CF is now cosine rather than Pearson.
  • Item-item models can now have unbounded size and have rows truncated by threshold value rather than size (#171).
  • FunkSVD now incorporates user ratings at predict time (#174).

Component Refactoring

This release contains a lot of algorithm and component refactorings. The public recommender API is unchanged, but many components and their interfaces have changed. This means that custom components and recommender configurations will need to be updated.

  • Added @Shareable annotation for components that can be put in the model build and shared between recommender sessions. All shareable components must now be annotated with this annotation (or configured via instance bindings), or they will not be pre-built.
  • Removed MinRating and MaxRating (deprecated in release 0.10). Code should use PreferenceDomain instead. See #163.
  • Renamed RatingSnapshot and related classes to PreferenceSnapshot, since that's what it really is.
  • Removed many annotations formerly used as dependency qualifiers (see #168).
    • @NormalizedSnapshot: components that formerly used this qualifier should depend on userNormalizedPreferenceSnapshot or use the normalizer directly.
    • @NormalizerBaseline: use a context-sensitive binding on BaselineSubtractingNormalizer.
    • @PredictNormalizer: use context-sensitive bindings. Usually this will require using context-sensitive bindings to bind the other normalizers.
    • @UserHistorySummary: was redundant with the component interface.
    • @UserVectorNormalizer: refactored into a separate UserVectorNormalizer interface, the default implementation of which delegates to a generic vector normalizer. Context-sensitive bindings configure the inner normalizer. BaselineSubtractingNormalizer has become BaselineSubtractingUserVectorNormalizer and implements the new interface.
    • @ItemSimilarity: factored into ItemSimilarity interface, with default implementation delegating to VectorSimilarity.
    • @UserSimilarity: factored into UserSimilarity interface, with default implementation delegating to VectorSimilarity.
    • @WeightedSimilarity: removed in favor of context-sensitive bindings.
    • Dropped @ClampingFunction (and DoubleFunction) in favor of a ClampingFunction interface. The new interface provides the user and item IDs, too.
  • Renamed HistorySummarizer to UserHistorySummarizer, and renamed base classes.
  • Moved normalizers and quantizers into subpackages of o.g.l.transform.
  • Refactored similarities for k-NN recommenders. Similarity became VectorSimilarity, without a generic parameter, and OptimizableVectorSimilarity and SymmetricBinaryFunction have been replaced by methods on similarity interfaces.
  • Removed ItemVector and UserVector classes. User and item IDs are now carried independently and we use distinct component interfaces to handle users and items.