Daniel Pyrathon - A practical guide to Singular Value Decomposition in Python

Recommender systems have become increasingly popular in recent years, and are used by some of the largest websites in the world to predict the likelihood of a user taking an action on an item. In the world of Netflix, this means recommending similar movies to the ones you have seen. In the world of dating, this means suggesting matches similar to people you already showed interest in!

My path to recommenders has been an unusual one: from a Software Engineer to working on matching algorithms at a dating company, with a little background on machine learning. With my knowledge of Python and the use of basic SVD (Singular Value Decomposition) frameworks, I was able to understand SVDs from a practical standpoint of what you can do with them, instead of focusing on the science.

In my talk, you will learn 2 practical ways of generating recommendations using SVDs: matrix factorization and item similarity. We will be learning the high-level components of SVD the "doer way": we will be implementing a simple movie recommendation engine with the help of Jupiter notebooks, the MovieLens database, and the Surprise recommendation package.

https://us.pycon.org/2018/schedule/presentation/96/

Transcript

1. A Practical Guide to SVD
2. Have you ever wondered why you binge-watch N E T F L I X ?
3. How did Spotify find out you secret love for Taylor Swift?
4. Recommendations
5. Who am I? • • •
6. What is a Recommendation Engine?
7. Why are recommendation engines important? 35% of Amazon’s revenue are generated by its recommendation engine 75% of users select movie based on Netflix’s recommendations
8. Collaborative Filtering
9. Collaborative Filtering: Users, Products, Ratings Users Ratings Products
10. What is Collaborative Filtering?
11. What is Collaborative Filtering?
12. What is Collaborative Filtering?
13. What is Collaborative Filtering?
14. Singular Value Decomposition
15. Collaborative Filtering through SVD
16. What are features?
17. What are latent features?
18. SVD Visualized User Product Matrix
19. SVD Visualized
20. Products matrix Prediction( ) = -2.0 * -0.5 = 3.52 Predicting new scores performing dot product + 6.3 * 0.4
21. Comparing based on Latent Features LF1 Users matrix
22. Demo