LoginSign Up
← Back to all sparks
R

recommenderlab

AI-ASSISTANTS
Velocity0.0

Lab for Developing and Testing Recommender Algorithms

github.com/mhahsler/recommenderlab ↗

recommenderlab added hybrid recommenders, then spent five years absorbing upstream churn.

recommender-systemscollaborative-filteringevaluationsparse-matricesmaintenance-mode
◆Current state
recommenderlab provides the rating matrix classes, recommender algorithms and evaluation schemes used to benchmark collaborative filtering in R. The algorithm surface has been settled since 0.2-6 added hybrid recommenders and 0.2-5 added a LIBMF-based one. Every release since has been reactive: sparse matrix coercion changes from Matrix, a cosine similarity fix from proxy, and most recently a dissimilarity change inherited from arules.
◆Where it's heading
The package sits on a stack it does not control — Matrix, proxy and arules — and the release notes read as a log of that stack moving. Three separate releases exist to track Matrix coercion and row/colSums changes alone. The genuine user-facing work now goes into evaluation ergonomics rather than algorithms: dropping users with too few ratings with a warning, making UBCF work when fewer than n neighbors exist, and accepting tibbles in coercion.
◆Prediction
The next release will most likely respond to another change in Matrix, proxy or arules, which have driven the last four. The 0 versus NA handling in sparse matrices flagged in 1.0-7 is the open thread most likely to need follow-up.

◆Recent moves

  1. 1y ago

    recommenderlab 1.0-7 accepts tibbles, tracks an arules change

    Coercion from data.frame to realRatingMatrix gains drop = TRUE so tibbles work, and sparse matrix handling distinguishes 0 from NA more carefully — a distinction that decides whether an unrated item is treated as a zero rating. The dissimilarity fix tracks an argument change in arules, continuing the pattern of upstream-driven releases.

    View source ↗
  2. 2y ago

    recommenderlab 1.0.5: interestMeasure and Matrix fixes

    A parameter rename in interestMeasure(), a fix for adding a single interest measure, and another Matrix row/colSums correction. Housekeeping against moving dependencies.

    View source ↗
  3. 3y ago

    recommenderlab 1.0.4 digest: evaluationScheme filtering and speed

    The substance is in the rolled-in 1.0.3 notes: evaluationScheme now drops users with too few ratings with a warning and builds faster for realRatingMatrix, and UBCF works for users with fewer than n nearest neighbors. Practical fixes for benchmarking on sparse real-world data.

    View source ↗
  4. 3y ago

    recommenderlab 1.0.2 digest: proxy cosine fix, Matrix prep

    Covers 1.0.0 through 1.0.2. Notable for requiring proxy >= 0.4-26 to pick up a cosine similarity conversion fix — the same upstream correction that parallelDist responded to in its own release, showing one dependency bug rippling across unrelated packages.

    View source ↗
  5. 5y ago

    recommenderlab 0.2-7 deprecates getConfusionMatrix for getResults

    Evaluation results get a clearer accessor, MovieLense gains user metadata, and a normalization bug in UBCF for binary data is fixed. The deprecation tidies an API that had grown around the evaluation workflow.

    View source ↗
  6. 6y ago

    recommenderlab 0.2-6 adds hybrid recommenders

    ⚡ SPARK

    The last release to add a genuinely new kind of model: HYBRID composes several recommenders into one, and crucially works inside evaluate(), so combinations can be benchmarked on the same footing as individual algorithms. Everything after this release is maintenance against upstream packages.

    View source ↗