Build the Movie Recommender

Recommender

Part of: Embeddings & Semantic Search

The "Because you watched Die Hard..." row on every streaming service is cosine similarity over movie vectors, and you now have every piece to build it. What a recommender is A recommender ranks items by similarity to something a user already likes. There is no magic: it turns "what should I watch next?" into "which vectors point closest to the vector of what you just watched?" The seed is the item you are finding neighbors for, and the answer is the nearest few vectors. How it works The recipe is four steps: 1. Give each movie a vector. Real systems learn these from viewing data; here we hand-craft genre scores so you can read them. 2. To recommend for a movie, compute cosine similarity from its vector to every other movie. 3. Sort the results descending. 4. Take the top N, that is the list. Our vectors have five dimensions: [action, romance, comedy, scifi, drama], each from 0 to 1. Die Hard is heavy action, light everything else. Interstellar leans sci-fi and drama. The math surfaces the right neighbors. That if name != title is not optional. A movie is always perfectly similar to itself (cosine 1.0), so recommending Die Hard to a Die Hard fan would top the list while being useles

Challenge: Recommend From a Taste Profile