Making Money From Movies - Step 4, Part 1

Step 4: Machine Learning Models

This is the fourth entry in a series about using machine learning models to predict movie success. You can find the first entry here. In previous posts, I discussed how I cleaned the data as well as did exploratory data analysis. The aim of this series is to answer the following questions:

1. Is there a correlation between genre and profits?
2. Is there a correlation between content rating and profits?
3. Are some genres more profitable for specific content ratings?
4. Does an increase in the number of Facebook likes for directors or actors correlate with increased profits?
5. Is being in or directing a greater number of movies correlated with more profits?
6. Are some topics correlated with more profits?

Linear Regression

Linear regression looks for the ability to predict a dependent variable from one independent variable or multiple variables. I chose to use it in this analysis because it is a quick way to see if a direct correlation exists between any features and profits. I first looked at the correlation between individual attributes against profits to better understand how strongly correlated each variable was with increased profits. I then combined features that appeared to correlate with increased profits to see if I could create a model that could predict earnings given new movies that the model had not yet seen. I used the normalized version of profits in these analyses because I found it easier to interpret the coefficient of determination using this value. The coefficient of determination, also called R squared, measures how much of an effect the independent variables have on the dependent variable. Using the normalized version of the profits caused this value to remain between 0 and 1 in most cases. Values closer to one indicate that more of the profit value can be attributed to the independent variables. Values nearer to 0 suggest that the independent variables have little effect on profit.

First, I investigated the relationship between genre and profits. I compared all the genres in the data set. Then I used a smaller set with just those genres that appeared to be more closely linked to a movie either losing money or being very profitable. The genres that were used were: ‘Thriller’, ‘Fantasy’, ‘Animation’, ‘Family’, ‘Action’,’ Drama’, ‘Sci-Fi’, ‘Crime’, ‘Adventure’ and ‘Romance’. The comparison with all of the genres gave slightly better results.

Next, I performed linear regression with the G, PG, PG-13 and R content ratings.

Looking at the coefficient of determination, it seems that very little of the variation in profit can be attributed to the content rating.

This is not the case for the number of movies that directors and actors have taken part in before working in a particular film which shows a slightly positive correlation. However, it is not large enough that it would be the primary recommendation for increasing profits. One reason for this might be that salaries presumably increase with the number of past movies an actor or director has been part of.

Facebook likes were the next feature that I tried. Facebook has not been around for as long as many of the movies in the database. In addition, its use has varied over time, so I looked at how Facebook likes predicted profits in movies made only after particular dates.

)

With films made after 2005, shortly after Facebook was founded, the coefficient of determination is only 9%. However, this increases to 17% with movies made after 2015. Facebook likes do a much better job of predicting success with films made more recently. Because the dataset contains movies from a wide range of dates, this feature might not be that useful in creating a model using this particular dataset. However, it would definitely be worth considering using this feature in a model with a set based off of only more recent movies.

Next, I looked at how well the plot keywords could predict movie profits. The graph below shows that these keywords have an abysmal predictive ability. Even though I used the normalized profits values, the coefficient of determination is way outside the range of 0 - 1. I tried a range of words and got very similar results each time.

Finally, I tried a model with the most predictive features from the previous models. I tried adding and removing various features. The best model I was able to make used:

• These content ratings: PG, G, R, PG-13 • These genres: Thriller, Fantasy, Animation, Family, Action, Drama, Sci-Fi, Crime, Adventure, Romance • The counts of the number of movies that the director and actors were in • These additional features: duration, facenumber_in_poster, genre_count

The combined features produced better results than any of the individual sets that I had tried. However, the coefficient of determination is pretty low, and I would not use this model to decide which movies to invest in.

Since the goal is to determine what movie producers can do to make a successful movie, it will be essential to see exactly which variables are more likely to lead to success. Using linear regression provided some good clues about which features influence movie success. Genre, number of movies a person was in, duration and ‘face_number_in_poster’ are likely to have some influence on profits. However, it doesn’t tell exactly which genres lead to greater profits or show a connection between particular genres and content ratings. To find that out, I will be using a decision tree algorithm and looking at the textual representation of the tree produced. You can read about it here.

Here is the code I used for the linear regression analysis.