Decompose a (sparse) user-item matrix to lower-rank matrices with f latent features. Includes naive user and item bias and regularization to limit overfitting. Constants untuned. Assumes a rating value between 1-5.
Model trained with stochastic gradient descent. Alternating least squares method pending.
##References
- Recommender Systems Handbook (Ch.5); Yehuda Koren and Robert Bell
- Matrix Factorization Techniques for Recommender Systems; Yehuda Koren, Robert Bell, and Chris Volinsky
- Netflix Update: Try This at Home; Simon Funk
- Probabilistic Matrix Factorization; Ruslan Salakhutdinov and Andriy Mnih
- Bayesian Probabilistic Matrix Factorization using Markov Chain Monte Carlo; Ruslan Salakhutdinov and Andriy Mnih