Below, we load the packages required for data analysis.
library(knitr)
library(tidyverse)Below, we load the movie ratings data from which we will build our recommender system and store it in a data frame.
my_url <- "https://raw.githubusercontent.com/geedoubledee/data607_week11/main/Week11_MovieRatings.csv"
movie_ratings_df <- read_csv(my_url)
fix_col_names <- function(s){
s <- gsub("[()]", "", s)
s <- gsub(" ", "_", s)
s
}
colnames(movie_ratings_df) <- map_chr(colnames(movie_ratings_df), fix_col_names)
movie_ratings_df <- movie_ratings_df %>%
column_to_rownames(., "User")We create two additional data frames from our data, one to capture the mean ratings by user:
baseline <- round(mean(as.matrix(movie_ratings_df), na.rm = TRUE), 2)
movie_cols <- c("M3GAN_2023", "The_Menu_2022", "Barbarian_2022", "Glass_Onion_2022", "Tár_2022", "Aftersun_2022")
movie_ratings_df_by_user <- movie_ratings_df %>%
mutate(User_Mean = round(rowMeans(.[, movie_cols], na.rm = TRUE), 2),
User_Relative_to_Baseline = User_Mean - baseline)
movie_ratings_df_by_user$User_Mean[is.nan(
movie_ratings_df_by_user$User_Mean)] <- NA
movie_ratings_df_by_user$User_Relative_to_Baseline[is.nan(
movie_ratings_df_by_user$User_Relative_to_Baseline)] <- NA
kable(movie_ratings_df_by_user, format = "simple")| M3GAN_2023 | The_Menu_2022 | Barbarian_2022 | Glass_Onion_2022 | Tár_2022 | Aftersun_2022 | User_Mean | User_Relative_to_Baseline | |
|---|---|---|---|---|---|---|---|---|
| Glen | NA | 4 | 5 | 4 | 5 | 3 | 4.20 | 0.35 |
| Sebastian | NA | 4 | 5 | 3 | 4 | 2 | 3.60 | -0.25 |
| Alex | NA | NA | NA | NA | NA | NA | NA | NA |
| Victoria | NA | NA | NA | 3 | NA | NA | 3.00 | -0.85 |
| Javin | NA | 4 | 4 | 3 | NA | NA | 3.67 | -0.18 |
| Matt | NA | NA | NA | NA | NA | 4 | 4.00 | 0.15 |
| Anne | NA | 4 | NA | 3 | 5 | NA | 4.00 | 0.15 |
| Grifin | NA | NA | NA | NA | NA | NA | NA | NA |
| Claire | NA | NA | NA | 4 | NA | NA | 4.00 | 0.15 |
| Vicki | NA | 4 | 5 | 3 | NA | NA | 4.00 | 0.15 |
| Dan | NA | 2 | 5 | NA | NA | NA | 3.50 | -0.35 |
| Frankie | NA | NA | NA | 3 | 5 | NA | 4.00 | 0.15 |
And another to capture the mean ratings by movie:
movie_ratings_df_by_movie <- as.data.frame(t(movie_ratings_df))
user_cols <- c("Glen", "Sebastian", "Alex", "Victoria", "Javin", "Matt",
"Anne", "Grifin", "Claire", "Vicki", "Dan", "Frankie")
movie_ratings_df_by_movie <- movie_ratings_df_by_movie %>%
mutate(Movie_Mean = round(rowMeans(.[, user_cols], na.rm = TRUE), 2),
Movie_Relative_to_Baseline = Movie_Mean - baseline)
movie_ratings_df_by_movie$Movie_Mean[is.nan(
movie_ratings_df_by_movie$Movie_Mean)] <- NA
movie_ratings_df_by_movie$Movie_Relative_to_Baseline[is.nan(
movie_ratings_df_by_movie$Movie_Relative_to_Baseline)] <- NA
kable(movie_ratings_df_by_movie, format = "simple")| Glen | Sebastian | Alex | Victoria | Javin | Matt | Anne | Grifin | Claire | Vicki | Dan | Frankie | Movie_Mean | Movie_Relative_to_Baseline | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| M3GAN_2023 | NA | NA | NA | NA | NA | NA | NA | NA | NA | NA | NA | NA | NA | NA |
| The_Menu_2022 | 4 | 4 | NA | NA | 4 | NA | 4 | NA | NA | 4 | 2 | NA | 3.67 | -0.18 |
| Barbarian_2022 | 5 | 5 | NA | NA | 4 | NA | NA | NA | NA | 5 | 5 | NA | 4.80 | 0.95 |
| Glass_Onion_2022 | 4 | 3 | NA | 3 | 3 | NA | 3 | NA | 4 | 3 | NA | 3 | 3.25 | -0.60 |
| Tár_2022 | 5 | 4 | NA | NA | NA | NA | 5 | NA | NA | NA | NA | 5 | 4.75 | 0.90 |
| Aftersun_2022 | 3 | 2 | NA | NA | NA | 4 | NA | NA | NA | NA | NA | NA | 3.00 | -0.85 |
Now we are ready to begin building our recommender system.
We set the initial predicted rating for each movie a user has not yet seen/rated as 0, and we set the initial predicted rating of movies they have already seen/rated as NA. This will prevent users from being recommended movies they’ve already watched.
recommendations_df <- movie_ratings_df %>%
replace(is.na(.), 0)
recommendations_df[recommendations_df > 0] <- NA
kable(recommendations_df, format = "simple")| M3GAN_2023 | The_Menu_2022 | Barbarian_2022 | Glass_Onion_2022 | Tár_2022 | Aftersun_2022 | |
|---|---|---|---|---|---|---|
| Glen | 0 | NA | NA | NA | NA | NA |
| Sebastian | 0 | NA | NA | NA | NA | NA |
| Alex | 0 | 0 | 0 | 0 | 0 | 0 |
| Victoria | 0 | 0 | 0 | NA | 0 | 0 |
| Javin | 0 | NA | NA | NA | 0 | 0 |
| Matt | 0 | 0 | 0 | 0 | 0 | NA |
| Anne | 0 | NA | 0 | NA | NA | 0 |
| Grifin | 0 | 0 | 0 | 0 | 0 | 0 |
| Claire | 0 | 0 | 0 | NA | 0 | 0 |
| Vicki | 0 | NA | NA | NA | 0 | 0 |
| Dan | 0 | NA | NA | 0 | 0 | 0 |
| Frankie | 0 | 0 | 0 | NA | NA | 0 |
Next, we calculate predicted ratings for all the movies a user has not yet seen/rated. The baseline rating will be the average rating across all movies, and each movie’s predicted rating for each user will be the sum of the baseline, the difference in that movie’s average rating relative to the baseline, and the difference in that user’s average rating relative to the baseline.
movie_relative_to_baseline <- movie_ratings_df_by_movie$Movie_Relative_to_Baseline
user_relative_to_baseline <- movie_ratings_df_by_user$User_Relative_to_Baseline
for (i in 1:nrow(recommendations_df)){
for (j in 1:ncol(recommendations_df)){
if (is.na(recommendations_df[i, j])){
next
}
else if (recommendations_df[i, j] == 0){
recommendations_df[i, j] <- sum(baseline,
movie_relative_to_baseline[j],
user_relative_to_baseline[i],
na.rm = TRUE)
}
}
}
kable(recommendations_df, format = "simple")| M3GAN_2023 | The_Menu_2022 | Barbarian_2022 | Glass_Onion_2022 | Tár_2022 | Aftersun_2022 | |
|---|---|---|---|---|---|---|
| Glen | 4.20 | NA | NA | NA | NA | NA |
| Sebastian | 3.60 | NA | NA | NA | NA | NA |
| Alex | 3.85 | 3.67 | 4.80 | 3.25 | 4.75 | 3.00 |
| Victoria | 3.00 | 2.82 | 3.95 | NA | 3.90 | 2.15 |
| Javin | 3.67 | NA | NA | NA | 4.57 | 2.82 |
| Matt | 4.00 | 3.82 | 4.95 | 3.40 | 4.90 | NA |
| Anne | 4.00 | NA | 4.95 | NA | NA | 3.15 |
| Grifin | 3.85 | 3.67 | 4.80 | 3.25 | 4.75 | 3.00 |
| Claire | 4.00 | 3.82 | 4.95 | NA | 4.90 | 3.15 |
| Vicki | 4.00 | NA | NA | NA | 4.90 | 3.15 |
| Dan | 3.50 | NA | NA | 2.90 | 4.40 | 2.65 |
| Frankie | 4.00 | 3.82 | 4.95 | NA | NA | 3.15 |
Finally, we can recommend the movie each user should watch next based on our expectations of them liking it:
recommendations_df$Recommendation <- colnames(recommendations_df)[apply(
recommendations_df, 1, which.max)]
Recommendation <- recommendations_df$Recommendation
names(Recommendation) <- user_cols
Recommendation <- as.data.frame(Recommendation)
kable(Recommendation, format ="simple", row.names = TRUE)| Recommendation | |
|---|---|
| Glen | M3GAN_2023 |
| Sebastian | M3GAN_2023 |
| Alex | Barbarian_2022 |
| Victoria | Barbarian_2022 |
| Javin | Tár_2022 |
| Matt | Barbarian_2022 |
| Anne | Barbarian_2022 |
| Grifin | Barbarian_2022 |
| Claire | Barbarian_2022 |
| Vicki | Tár_2022 |
| Dan | Tár_2022 |
| Frankie | Barbarian_2022 |
Because Glen and Sebastian had seen most of the movies on the list already, M3GAN was the only movie left to recommend to them.
All the other users were recommended to watch either Barbarian or Tár. This is not surprising given that, of all the movies in our data set, the average ratings for these two movies were nearly a full point higher than the average movie rating.
This recommender system could be enhanced by adding a movie-to-movie table that gave similarity scores from one movie to a fixed number of other movies. Genre would most likely be a large part of that similarity score, but shared directors/actors could also factor in. That would especially help when trying to recommend new movies that few users have seen/rated yet. It could also be enhanced by attempting to show users something they didn’t know they would like and bring them out of their movie silos. Generally, you only want to recommend those kinds of movies occasionally, so as not to make users feel like the algorithm doesn’t know them at all.