Getting started with recommenderlab
In recommenderlab: Lab for Developing and Testing Recommender Algorithms

knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
library(recommenderlab)
set.seed(1234)

recommenderlab provides tools for representing user–item data, fitting recommendation algorithms, producing recommendations, and evaluating their quality. This vignette walks through that workflow using the package's bundled MovieLense ratings data.

Installation

Install the released package from CRAN, then load it in your R session:

install.packages("recommenderlab")

The SVD and LIBMF recommenders require the optional packages irlba and recosystem, respectively. Install them if you plan to use those methods:

install.packages(c("irlba", "recosystem"))

library(recommenderlab)

Load and prepare ratings

The MovieLense data contains ratings on a one-to-five-star scale. It is stored as a sparse realRatingMatrix: users are rows, movies are columns, and missing ratings are not stored as zeros. We select users who rated more than 100 movies to give the recommendation algorithms enough information to work with.

data("MovieLense")
MovieLense

MovieLense100 <- MovieLense[rowCounts(MovieLense) > 100, ]
MovieLense100

Basic summaries help describe the data before modeling. For example, rowCounts() counts ratings per user, and getRatings() extracts the observed rating values.

summary(rowCounts(MovieLense100))
summary(getRatings(MovieLense100))

Fit a recommender and make recommendations

Recommender() learns a model from a training rating matrix. Here we fit user-based collaborative filtering (UBCF) on the first 300 selected users. predict() then produces a top-five list for two other users. The default output is a topNList; coerce it to a list to see the recommended movie titles.

train <- MovieLense100[1:300, ]
rec <- Recommender(train, method = "UBCF")
rec

recommendations <- predict(rec, MovieLense100[301:302, ], n = 5)
recommendations
as(recommendations, "list")

The package also supports predicted ratings. Request type = "ratings" when the numeric estimates are more useful than a ranked list.

predicted_ratings <- predict(
  rec,
  MovieLense100[301:302, ],
  type = "ratings"
)
as(predicted_ratings, "matrix")[, 1:6]

Evaluate recommendations

Evaluation should simulate the information available when recommendations are made. An all-but-five scheme withholds five ratings per user and uses the remaining ratings as known input. Here, ratings of four stars or higher count as positive feedback. evaluationScheme() supports train/test splits, cross-validation, and bootstrap evaluation.

evaluation_data <- MovieLense100[1:200, ]
scheme <- evaluationScheme(
  evaluation_data,
  method = "cross-validation",
  k = 5,
  given = -5,
  goodRating = 4
)
scheme

Compare a popularity-based recommender with a random baseline. evaluate() fits each method on every training fold, creates top-N recommendations, and calculates measures from the withheld ratings. The resulting true-positive and false-positive rates can be plotted to compare recommendation list lengths.

algorithms <- list(
  `popular items` = list(name = "POPULAR", param = NULL),
  `random items` = list(name = "RANDOM", param = NULL)
)

results <- evaluate(
  scheme,
  algorithms,
  type = "topNList",
  n = c(1, 3, 5, 10),
  progress = FALSE
)
getResults(results[[1]])

Plot the average true-positive rate against the false-positive rate for each recommendation list length:

plot(results, annotate = TRUE, legend = "topleft")

For predicted ratings, evaluate() can instead report rating error measures such as RMSE, MSE, and MAE by using type = "ratings". See ?calcPredictionAccuracy for the measures available for direct predictions and ?evaluate for details on evaluation results.

Where to go next

The package includes additional algorithms such as item-based collaborative filtering (IBCF), matrix factorization, association-rule recommenders, and hybrid recommenders. Use recommenderRegistry$get_entry_names() to see the methods available in your installation. The reference manual documents each algorithm, data class, and evaluation helper.