Blog

I've been thinking a lot about movies this summer. And I always come back to how can I add this or that to my recommendation formula to make it more accurate. Last week, I had collected enough improvements and decided it was time to implement them!

The first big change that I thought of was how similar movies are used to generate recommendations. I kept thinking, it's not just how similar a movie is, there's also generally a quality difference. Take for example the Matrix and the Matrix Revolutions. Obviously very similar movies, but users who have rated both movies rated the first movie an average of 1.3 higher than the third movie in the franchise. So if I give the Matrix an 8, I'll probably enjoy the Matrix Revolutions more around a 6 or 7 than an 8 as well. Of course this is not always the case, but when you do this for many movies the prediction on average becomes more accurate.

The second big change came from thinking about horror movies. I felt that my recommendations were overestimating how much I would enjoy horror movies. A large part of the reason was because one of the users with a similar taste in movies really likes horror movies. Whereas I generally don't. We match so well on comedy, drama, and romance that it compensated for the mismatch in horror. Which left me with inflated horror movie recomendation. So I decided to also calculate similarity based on genre for each user and only use the genres with a high similarity. Now this user is helping me with comedy, drama, and romance recommendations but not horror movies.

The third small change I made was making who counts as a user with a similar taste in movies slightly stricter. Specifically, a big difference in rating for a movie (for example, I rate a movie as a 9 and another users rates it as a 4) reduces the similarity to a greater extent. This will reduce the amount of big mistakes that occur in the recommendations.

Here are some of my new recommendations. Looking forward to watching them :)

Curious about your new and improved recommendations? Go check them out!