# Going to the Movies…

October 22, 2012
By

(This article was first published on Rcrastinate, and kindly contributed to R-bloggers)

Today, let us have a look at movies. The Internet Movie Database (IMDb) has some data dumps available on their website. It's a subset of the information available on the IMDb site, but it's more than enough. I will spare you my code to convert these data dumps in R dataframes, because the code is boring and complicated (unfortunately, the data dumps are not too nice to read automatically).

I just wanna show you what you can do with these dumps. I gonna use the data from IMDb user ratings (saved in the variable rat). A compressed .Rdata file of these ratings is roughly 7 MB big and has almost 390,000 rows. rat look like this:

> rat[grep("The Wire", rat$title),][1:10,] n.votes rating title year 25026 300 8.0 "Curb Your Enthusiasm" (2000) {The Wire (#1.6)} 2000 103831 644 8.0 "Star Trek: Deep Space Nine" (1993) {The Wire (#2.22)} 1993 131481 93 8.0 "The Wire" (1997) 1997 131482 2457 9.5 "The Wire" (2002) 2002 131483 843 9.5 "The Wire" (2002) {-30- (#5.10)} 2002 131484 213 9.2 "The Wire" (2002) {A New Day (#4.11)} 2002 131485 227 8.7 "The Wire" (2002) {All Due Respect (#3.2)} 2002 131486 250 8.9 "The Wire" (2002) {All Prologue (#2.6)} 2002 131487 553 8.8 "The Wire" (2002) {Alliances (#4.5)} 2002 131488 218 8.8 "The Wire" (2002) {Back Burners (#3.7)} 2002 series.ep ep.code season episode decade 25026 TRUE 1.6 1 6 1990 103831 TRUE 2.22 2 22 1990 131481 FALSE <NA> NA NA 1990 131482 FALSE <NA> NA NA 2000 131483 TRUE 5.10 5 10 2000 131484 TRUE 4.11 4 11 2000 131485 TRUE 3.2 3 2 2000 131486 TRUE 2.6 2 6 2000 131487 TRUE 4.5 4 5 2000 131488 TRUE 3.7 3 7 2000 First, let's have a look at the overall distribution of all ratings. Let's exclude episode ratings, we are only interested in movies and whole series right now. Let us start with a nice histogram. library(MASS) truehist(rat[rat$series.ep == F, "rating"], border = "#00000000", col = "darkblue", xlab = "Rating", ylab = "Probability")
abline(v = mean(rat[rat$series.ep == F, "rating"]), col = "lightgreen", lwd = 2) curve(dnorm(x, mean = mean(rat[rat$series.ep == F, "rating"]), sd = sd(rat[rat$series.ep == F, "rating"])), from = 1, to = 10, add = T, lwd = 2, col = "red", lty = "dotted") rat.tab <- table(rat[rat$series.ep == F, "rating"])
rat.tab[which(rat.tab == max(rat.tab))]

The mean movie rating (green line) on IMDb is 6.14 (rounded). Users can rate movies on IMDb between 1 and 10. Movie ratings on IMDb are not normally distributed but slightly shifted to the right. A normal distribution with the same mean and standard deviation as the ratings is included in the histogram, it's the dotted red line.

Since we have information about the decade a movie was published, let's have a look at ratings over the decades. Note, the year noted at the left of the plot is the start of that decade.

dotplot(rev(xtabs(rating ~ decade, data = rat[rat$series.ep == F,]) / xtabs(~ decade, data = rat[rat$series.ep == F,])), cex = 1.3, xlab = "Mean rating")
Wow, that's harsh - movies obviously sucked from 1900 to 1909. The "Golden Twenties" win with a mean rating of 6.40. That's very close to movies from the 2nd best decade (1940 to 1950) which have a mean rating of 6.35. I tried to create this plot with error bars to visualize the variance in decades. However, error bars are practically invisible because there are so many cases in each decade. The decade with the least ratings is 1910 to 1919 with "only" 1,854 ratings. The decade with the most ratings is the one between 2000 and 2009 - it has 108,304 ratings! It's no surprise that error bars are practically invisible with such high counts.

Now we gonna look into series. There is always a huge discussion going on which season of a series is the best. Let's have a look what IMDb users say. I gonna plot the mean of each season with error bars to get an impression of statistical significance. Note, that I gonna plot the mean of means because the data dump you can download from IMDb only supplies mean ratings of episodes. Normally, I would calculate the mean for each season based on "raw" user ratings.

I gonna compare three series which many people say they were the best they ever saw.

First: Define some patterns you want to find in the data dump.
"\"The Wire\"",
"\"The Sopranos\"")

Now build up a dataframe with all hits.
rat.series <- data.frame()
for (series in series.patterns) {
rat.series <- rbind(rat.series, rat[grep(series, rat$title, fixed = T),]) } Now, I'm extracting the series title (I only need that for the legend of the plot). rat.series <- rat.series[rat.series$series.ep == T,]
rat.series$series.title <- sapply(rat.series$title, USE.NAMES = F, FUN = function (title) {
ti <- grep("[\"]{1}[[:print:]]*[\"]{1}", strsplit(title, "(", fixed = T)[[1]], value = T)
gsub("[\"[:space:]]", "", ti) } )

Now plot the result with the help of a function from the "sciplot" package.
lineplot.CI(season, rating, series.title, data = rat.series, col = c("#FF0000C8", "#00FF00C8", "#0000FFC8"), lwd = 2, xlab = "Season", ylab = "Mean Rating")
For me being such a big fan of "The Wire" that's a tough result. IMDb users say that "Breaking Bad" is at least equally outstanding great. I just started watching "Breaking Bad", so I guess it's alright. "The Sopranos" seem to suck after season 3... I quit within season 3. Maybe, that's alright, too :)

A second duel is between two cartoon classics.
Both are rated worse over time with a few peaks at season 8 and 11 (Imaginationland?!? Come on, people, you can't be serious!) for South Park.

If you have any ideas for other duels, let me know.

But now to the answer we've all been waiting for: The best series EVER (at least in the eyes of voting IMDb users). Place your bets, ladies and gentlemen.

rat.se is a dataframe holding all data for episodes (so no movie ratings). We only want to look at the 250 series with the most votes. So we cross-tab number of votes over the title of the series (I already showed above how I used regular expressions extracting the title of the series). We also sort the xtab and take the first 250 entries.

Now I gonna use the names of xt.votes, iterate through them to calculate means, standard errors and number of votes for each series. The result is saved in the variable df. I sort this dataframe by rating and extract the Top 10.
df <- data.frame()
mean.rat.se <- mean(rat.se.tt[rat.se.tt$series.title == name, "rating"]) se.rat.se <- se(rat.se.tt[rat.se.tt$series.title == name, "rating"])
nvotes.rat.se <- sum(rat.se.tt[rat.se.tt$series.title == name, "n.votes"]) new.df <- data.frame(title = name, mRating = mean.rat.se, SERating = se.rat.se, nVotes = nvotes.rat.se) df <- rbind(df, new.df) } df <- df[order(df$mRating, decreasing = T),][1:10,]

Aaaaand ... plot!
par(mar = c(10, 4, 4, 2))
plot(df$mRating, axes = F, ylab = "Mean Episode Rating", xlab = "", pch = 19, col = "blue", cex = 1.5, ylim = c(8.4,8.9), xlim = c(0.7,10.3)) plotCI(x = df$mRating, ui = df$mRating + df$SERating, li = df$mRating - df$SERating, add = T, type = "n", col = "darkblue")
axis(side = 2)
axis(side = 1, labels = df\$title, at = 1:10, tick = F, las = 2, cex.axis = 0.8)

Click the plot to read the axis labels...

So we have a winner: Game of Thrones! The error bars represent standard errors of the mean ratings for each episode - that's not the waterproof way to do it (wo would need the raw ratings for that). Nevertheless, we can see some interesting things:
• There are only very small differences between the Top 10 series. Look at the scale of the y axis. The difference between Game of Thrones and The Wire is really really small!
• "Firefly" and "Sherlock" have relatively great error bars. The reason: There are only few episodes of them. "Sherlock" still has a chance to gather more episodes. Unfortunately, Firefly does not.
• Dexter is overrated ;)
Enough of the movies and series for now... Maybe, we will come back to that dataset some other time.