Factorization machines
Ref: https://securityintelligence.com/factorization-machines-a-new-way-of-looking-at-machine-learning/
A new way of looking at machine learning
- Introduced by Steffen Rendle 2010 at Google.
- Can be compared to SVM with a polynomial kernel.
- SVM is kind of mystery(?)
- SVM works best on dense data.
- Factorization machines perform well on sparse data.
- Can model \(n\)-way variable interactions.
- Computational complexity can be reduced to linear.
- Optimization methods: stochastic gradient descent, alternating least-squares, Markov chain Monte Carlo (recommended), adaptive gradient descent.
- Widely used in collaborative recommendation systems.
- Recommend music, movies; predict stock market.
-
libFM, fastFM, spark-libFM