DATA 612 Project 5 | Implementing a Recommender System on Spark

Introduction

The goal of this assignment is to adapt one of my recommendation systems to work with Apache Spark and compare the performance with my previous iteration. I will adapt my MovieLense recommendation systems from project 2 and compare it with the model built on Apache Spark.

# load libraries
library(tidyverse)
library(kableExtra)
library(knitr)
library(recommenderlab)
library(sparklyr)

Dataset

I will use the MovieLense dataset; the 100k MovieLense ratings data set. The data was collected through the MovieLens web site (movielens.umn.edu) during the seven-month period from September 19th, 1997 through April 22nd, 1998. The data set contains about 100,000 ratings (1-5) from 943 users on 1664 movies.

# load dataset
data("MovieLense")

Centralized Recommender System

For project 2, I worked with MovieLense dataset to build item-based collaborative filter and user-based collaborative filter models. However, sparklyr library does not offer such algorithms. Therefore, for comparison, we will implement ALS method for both recommender systems.

I will use the k-fold method to split the data. Then fit the ALS model and make prediction. Lastly, I will extract accuracy of the model to compare with the model built on Spark.

# start 
a1 <- Sys.time()

# split dataset using k-fold method
set.seed(156)
scheme <- MovieLense %>% 
  evaluationScheme(method = "cross", k = 5, given = 15, goodRating = 3)

# fit the model
als_model <- Recommender(getData(scheme, "train"), method = "ALS")

# make prediction
prediction <- predict(als_model, getData(scheme, "known"), type = "ratings")

# end 
z1 <- Sys.time()

# get accuracy score
evaluation <- calcPredictionAccuracy(prediction, getData(scheme, "unknown"))

Distributed Recommender System

In this section I will adapt the model above to work with Apache Spark. Apache Spark is open source, general-purpose distributed computing engine used for processing and analyzing a large amount of data.

I will create a connection to spark and then copy the dataset to spark. Sparklyr package accepts only data frame and numeric variables so I will make the changes. Next, I split the dataset in the spark platform and fit the model. Lastly I will make prediction and copy the result to R.

# convert data based on sparklyr requirements
sdf_MovieLense <- MovieLense %>% 
  as(. , "data.frame") %>% 
  mutate(user = as.numeric(user),
         item = as.numeric(item)) 

# connect to spark locally
sc <- spark_connect(master = "local")

# start 
a2 <- Sys.time()

# copy data to spark
sdf_rating_matrix <- sdf_copy_to(sc, sdf_MovieLense, "sdf_rating_matrix", overwrite = TRUE)

# split dataset in spark
partitioned <- sdf_rating_matrix %>% 
  sdf_random_split(training = 0.8, testing = 0.2)

# fit the model
sdf_als_model <- ml_als(partitioned$training, max_iter = 5)

# make prediction
prediction <- ml_transform(sdf_als_model, partitioned$testing) %>% collect()

# end 
z2 <- Sys.time()

# disconnect from spark
spark_disconnect(sc)

## NULL

Performance Evaluation

To evaluate the performances of the models, I will compare the RMSE of the models and the time it took from data preparation to data prediction for both models.

# function to calculate RMSE
rmse <- function(o, p) {
  round((sqrt(mean((o - p)^2, na.rm = TRUE))), 2)
}

# rmse for both models
rmse1 <-  evaluation[[1]]
rmse2 <- rmse(prediction$rating, prediction$prediction)

# print the score
kable(cbind(rmse1, rmse2), col.names = c("recommenderlab", "sparklyr")) %>% 
  kable_styling("striped", full_width = F) %>% 
  add_header_above(c("RMSE" = 2))

# print the score
kable(cbind((z1 - a1), (z2 - a2)), col.names = c("recommenderlab", "sparklyr")) %>% 
  kable_styling("striped", full_width = F) %>% 
  add_header_above(c("Processing Time" = 2))  

Conclusion

For this assignment, I used a dataset of 100 thousand ratings and built two recommender systems. One on my personal computer and the other on distributed general-purpose cluster-computing framework. Even though I used the same dataset, I have an improved RMSE score for the model built on Spark. I assume it is due to the number of maximum iteration the model ran. (P.S. I could not run the default of 10 iteration, I got an error and could not find a solution.) I also compared the time it took for the model to fit and make prediction. As expected Spark did the calculation in less than half.

In conclusion, with this particular dataset, it is best to move to a distributed platform before making prediction because that is the only time my computer froze for several seconds.