Author(s): Arjun Bhasin

Originally published on Towards AI.

How Netflix guesses what you will watch. And why a huge table of ratings hides a few simple ‘tastes’.

Suppose that you finish a show on Netflix and, before the credits end, it suggests another one that you actually want to watch, even though you never told Netflix what kind of movies you like. How does it make that guess?

After the introduction, the article explains that Netflix’s problem can be viewed as filling in a mostly empty “ratings matrix,” where the key insight is that people’s preferences are driven by only a few underlying tastes. It walks through a simplified example (friends with ratings across a handful of movie types) to show how recommendations arise from representing each person and movie with short vectors and predicting ratings via a dot product (low-rank structure). The discussion then connects this idea to singular value decomposition (SVD), illustrating how a complete table can be reconstructed from a small number of dominant layers (tastes) while smaller layers mostly add noise. Because real data has many missing cells, the article describes the “Funk SVD” approach popularized during the Netflix Prize: instead of doing full SVD, it directly learns the two low-dimensional “thin tables” by iteratively adjusting taste vectors to reduce error on known ratings while using a regularization penalty to avoid memorizing. Finally, once the blanks are effectively filled with predicted ratings, the system recommends the unseen movies with the highest predicted scores, and the article closes by noting practical additions like per-person/movie biases, behavioral signals beyond stars, and changing tastes over time.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI