A user doesn't just like a 'Genre'; they like a specific mix of humor, pacing, and mood. Matrix Factorization is the math that finds that specific recipe.
1The Hidden Features
Traditional recommenders use visible features like 'Genre' or 'Director'. Matrix Factorization discovers Latent Factors—hidden characteristics that the model identifies through the patterns of user behavior. One factor might represent 'Movies with dry humor and slow pacing,' and another 'Explosive blockbusters with simple plots.' The model doesn't give these factors names, but it assigns a numerical value to every user and item for each factor, creating a dense map of interest.
2SVD: The Scalpel
Singular Value Decomposition (SVD) is the mathematical engine that performs this split. It takes the sparse Rating Matrix and factorizes it into three parts: a user matrix, a diagonal matrix of 'Singular Values' (the importance of each factor), and an item matrix. By keeping only the top factors (dimensionality reduction), we remove the 'Noise' (random ratings) and focus on the 'Signals' (real user trends). This is what made SVD the dominant algorithm in the Netflix Prize competition.
3Reconstructing Reality
Once the matrices are factorized, predicting a rating is simple. To find out if User X will like Movie Y, we take the Dot Product of User X's factor vector and Movie Y's factor vector. If both have high weights for the same hidden factors, the resulting score will be high. This 'Compression' approach is extremely powerful for Sparse Data, as it allows the model to 'Infer' preferences even if a user has only rated a handful of items.
