Matrix Factorization with SVD
User-item matrix, SVD decomposition, latent factors, Surprise library implementation.
Why Matrix Factorization?
The user-item matrix is huge and mostly empty. Matrix factorization compresses it into two smaller matrices of latent factors, one for users and one for items, whose product reconstructs (and fills in) the ratings. This handles sparsity far better than neighbor methods.
Latent Factors Intuition
Each user and item is described by a short vector of hidden features (maybe action-ness or comedy-ness for movies). A predicted rating is the dot product of a user vector and an item vector, learned purely from the data.
All lessons in this course
- Collaborative Filtering: User-Based and Item-Based
- Matrix Factorization with SVD
- Content-Based Filtering
- Hybrid Systems and Evaluation Metrics