Matrix factorization decomposes the user-item matrix into latent factors.
Model: where U is users × factors, V is items × factors.
Learning: Minimize reconstruction error on known ratings. Regularize to prevent overfitting.
SVD variants: Truncated SVD, SVD++, probabilistic matrix factorization.
Why it works: Latent factors capture underlying preferences (action lover, art film fan) without explicit labels.
Interview question: "How handle implicit feedback (clicks, not ratings)?"
Weight by confidence. Treat non-interactions as weak negatives. Use ALS optimization.