What is SVD (Singular Value Decomposition) in recommendation systems?
Updated May 16, 2026
Short answer
SVD decomposes the user-item matrix into latent factors to predict missing ratings.
Deep explanation
SVD factorizes a sparse user-item matrix into three matrices (U, Σ, Vᵀ), capturing latent relationships between users and items. In recommender systems, it helps reduce dimensionality and predict missing interactions by reconstructing approximations of the original matrix. Unlike basic matrix factorization, SVD explicitly models orthogonal latent features and is often regularized to handle sparsity.
Real-world example
Netflix-style movie recommendation using latent user preferences for genres and actors.
Common mistakes
- Assuming SVD works directly on sparse matrices without preprocessing or handling missing values.
Follow-up questions
- How is SVD different from ALS?
- Why is SVD not directly used on large sparse data?