Movie Recommendations

Description

Use the MovieLens data to recommend movies to users.

Objectives

After completing this assignment, you should be able to:

Use the csv module to read files
Use list comprehensions to filter data and perform calculations
Model a problem using objects and functions

Details

Use via $ python3 movies.py The options will follow from there

Excepted files

This program expects the existence of directories

ml-100k
ml-1m

These directories should hold the unzipped contents of the MovieLens data.

Normal Mode

First, go to the MovieLens website and download the MovieLens 100K data. Unzip it and read the README file to understand the data.

Your goal is going to be to write a system that will recommend movies to a user.

Step 1

You need to be able to load in movie and rating data. Using the csv module, write a module that will load in the movie data from u.item and the rating data from u.info. You can choose how you will store movies and ratings, but you will need to be able to associate them later.

Specifically, you will need to be able to:

Find all ratings for a movie by id
Find the average rating for a movie by id
Find the name of a movie by id
Find all ratings for a user

Step 2

The easiest way to recommend movies is to recommend the most popular movies. Write a program to show the top X movies by average rating with their rating. You need to be able to state a minimum number of ratings for a movie to be considered.

Now, create the ability to find the top X movies by average rating that a specific user has not rated. This allows you to suggest popular movies for a specific user.

Step 3

Popular movies are not really good enough on their own. What would be great is a way to match two users by their tastes. You need to create the ability to take two users and find their similarity. There's a few ways to do this. We'll focus on the Euclidean distance. If you have a list of movie ratings for user 1 (v) and a list for user 2 (w), where each list is made up of ratings for movies they've both seen in the same order, then you can use this formula:

def euclidean_distance(v, w):
    """Given two lists, give the Euclidean distance between them on a scale
    of 0 to 1. 1 means the two lists are identical.
    """

    # Guard against empty lists.
    if len(v) is 0:
        return 0

    # Note that this is the same as vector subtraction.
    differences = [v[idx] - w[idx] for idx in range(len(v))]
    squares = [diff ** 2 for diff in differences]
    sum_of_squares = sum(squares)

    return 1 / (1 + math.sqrt(sum_of_squares))

You may want to look up the Pearson correlation score. This is more complicated, but accounts for people with different grading scales (for example, I may never rate movies above 4 because I am grumpy, but our relative scoring may be similar.)

Step 4

Now that you can calculate the similarity between two users, add a new ability. Given a list of all users, find the users most similar to a specific user, and then recommend the highest rated movies from those users that the specific user hasn't seen.

A good formula for figuring out movies that user might like the most is similarity * rating.

Step 5

Put this all together! The interface is up to you. You may want to have one program that presents a menu system so you can see top overall movies, popular movies you haven't seen (you'll have to give your user id), or recommendations specific to you.

Another option would be a command-line program that takes arguments on the command line. Look at the argparse library for this. You might make multiple programs, like so:

popular_movies.py -- returns a table of popular movies, takes a user_id argument to filter out movies that user has seen
recommendations.py -- returns a table of recommended movies for a user

Hard Mode

In addition to all of the above:

Look up the Pearson correlation score and implement it as well as Euclidean distance. Try both and find out which gives you better results.
We currently recommend movies by finding similar users. Turn your data around to find similar movies based on their reviews from users. Add the ability to choose a movie by id and see the movies that are most like it and most not like it.
Try out your program with the MovieLens 1M dataset.
Think of something new and try it out! There's a lot of data that comes with MovieLens.

AlanCoding/movie-recommendations