movie-rec-tags

Content-based movie recommender system using MovieLens tags

Environment YAML File

I included a conda YAML file for you to create the environment I used with all the libraries installed:

conda env create -f env_movie_rec_tags.yml

Motivation for the Project

I wanted to see if I can build a simple but reasonable content-based movie recommender system with only MovieLens tags. There are three key business questions that I answer along the way which are critical to building this recommender system.

Files

Movie-Rec-Movielens-Tags.ipynb
- Jupyterlab notebook with all the code used in the analysis
env_movie_rec_tags.yml
- YAML File with configuration of environment used

Data

All data used is from the MovieLens 20M datasets: https://grouplens.org/datasets/movielens/20m/

Medium Blog Post

https://medium.com/@johnson.h.kuan/how-to-build-a-simple-movie-recommender-system-with-tags-b9ab5cb3b616

Summary of Results

I was able to build a simple content-based movie recommender system with only MovieLens tags by answering three key business questions (see blog post for more details):

How many tags do we need for each movie?
- I kept the top N tags for each movie based on relevance score
- I demonstrated that N = 50 was reasonble to make sure we had enough relevant tags for each movie
How do we use tags to measure the similarity between movies?
- I demonstrate two appproaches that were reasonble:
  - Jaccard Distance of two sets of movie tags
  - Cosine Similarity of Movie Vectors (aka Content Embeddings) based on tags
How do we use tags to generate movie recommendations for a user?
- I demonstrated that we can compute a user vector based on the average of movie vectors that the user has watched
- I demonstrated that we can use this user vector to find similar movies based on cosine similarity
- These similar movies are the recommendations to the user
- I generated recommendations based on movies that I like
- The recommendations were suprisingly good considering that I only used MovieLens tags