HHernandez_DATA612_Project5

Recommender System based on Alternating Least Squares (ALS) - recommenderlab Implementation

tic()

# Build a user matrix with movies as columns

rating_mat <- spread(ratings[,1:3], movieId, rating)
rating_mat <- as.matrix(rating_mat[,-1]) #remove userIds

library(recommenderlab)

#Convert into a recommenderlab sparse matrix
rating_mat <- as(rating_mat, "realRatingMatrix") 

#Focus on a more relevant set of ratings with the following constraints: Minimum of 5 users per rated movie and 5 views per movie.

ratings_relevant <- rating_mat[rowCounts(rating_mat) > 5,            colCounts(rating_mat) > 5]

ratings_relevant

## 610 x 3268 rating matrix of class 'realRatingMatrix' with 88364 ratings.

# Creating Recommender Model (ALS) and Evaluating Model using Cross-validation

eval_sch <- evaluationScheme(rating_mat, method="cross-validation", k=10, given=3, goodRating=3)

recommender_model <- Recommender(getData(eval_sch,"train"), method = "ALS")

recom <- predict(recommender_model, newdata=getData(eval_sch,"known"), n=10, type="ratings") 

# Model Performance
eval_accuracy <- calcPredictionAccuracy(x = recom, data = getData(eval_sch, "unknown"), given=10, goodRating=3, byUser = FALSE)
eval_accuracy

##      RMSE       MSE       MAE 
## 1.0021374 1.0042793 0.7664364

rlab_time <- toc(quiet = FALSE)

## 278.06 sec elapsed

Recommender based on Alternating Least Squares (ALS) - Spark Implementation

tic()

# Spark connection

sc <- spark_connect(master = "local")

# Spark data preparation

sprk_ratings <- ratings[,-4]

train_filter <- sample(x = c(TRUE, FALSE), size = nrow(sprk_ratings),replace = TRUE, prob = c(0.8, 0.2))

train_ratings <- sprk_ratings[train_filter, ]
test_ratings <- sprk_ratings[!train_filter, ]

# Copy data to Spark memory
sprk_train <- sdf_copy_to(sc, train_ratings, "train_ratings", overwrite = TRUE)
sprk_test <- sdf_copy_to(sc, test_ratings, "test_ratings", overwrite = TRUE)


# Creating Recommender Model (ALS)
sprk_recommender_model <- ml_als(sprk_train, max_iter = 5, nonnegative = TRUE, 
                   rating_col = "rating", user_col = "userId", item_col = "movieId")

sprk_recom <- sprk_recommender_model$.jobj %>%
  invoke("transform", spark_dataframe(sprk_test)) %>%
  collect()

#Remove NAs to perform evaluation calculations
sprk_recom <- sprk_recom[!is.na(sprk_recom$prediction), ]

# Evaluating Model
sprk_mse <- mean((sprk_recom$rating - sprk_recom$prediction)^2)
sprk_rmse <- sqrt(sprk_mse)
sprk_mae <- mean(abs(sprk_recom$rating - sprk_recom$prediction))

sprk_mse

## [1] 0.7716416

sprk_rmse

## [1] 0.8784313

sprk_mae

## [1] 0.6782136

# Spark disconnection
spark_disconnect(sc)

## NULL

sprk_time <- toc(quiet = FALSE)

## 46.45 sec elapsed

# Comparison

comp_methods <- rbind(eval_accuracy, data.frame(RMSE = sprk_rmse, MSE = sprk_mse , MAE = sprk_mae))
rownames(comp_methods) <- c("recommenderlab", "Spark")
library(knitr)
kable(comp_methods)

Conclusion

Spark implementation of the ALS-based recommender model outperformed recommenderlab in both main areas: Performance and Accuracy.

Performance was expected as it is one of the main benefits of using a distributed data/analytics platform, although installed as Local Node, Spark uses multithreading to achieve the distributed compute fashion.

Spark outperformed recommenderlab by a factor of 6X in terms of performance

(R can also be run as a parallel engine with R Studio Server and MS R Open multithreaded execution).

Spark running in a multi-server and high-density environment, like in the cloud, can achieve unparalleled performance that makes possible addressing the most challenging DS and ML problems. Cloud elasticity allows to minimize compute and maintenance costs while maximizing efficiency.

The Spark implementation also showed better accuracy, 0.88 RMSE vs 1.10 RMSE in recommenderlab, I am thinking there is also a better ALS implementation in Spark but the exact reason is not 100% clear for me.