As author describes:
FMs combine the high-prediction accuracy of factorization models with the flexibility of feature engineering. The input data for FMs is described with real-valued features, exactly like in other machine-learning approaches such as linear regression, support vector machines, etc. However, the internal model of FMs uses factorized interactions between variables, and thus, it shares with other factorization models the high prediction quality in sparse settings, like in recommender systems. It has been shown that FMs can mimic most factorization models just by feature engineering [Rendle 2010]
Steffen Rendle (2012): Factorization Machines with libFM, in ACM Trans. Intell. Syst. Technol., 3(3), May. [PDF]
Implementation is mostly based on libfm software and try to be compatible with conventions used there.
using FactorizationMachines
T = [
5 1 0 1 0 0 0 1 0 0 12.5;
5 1 0 0 1 0 0 1 0 0 20;
4 1 0 0 0 1 0 1 0 0 78;
1 0 1 1 0 0 0 0 0 1 12.5;
1 0 1 0 1 0 0 0 0 1 20;
]
X = sparse(T[:,2:end])'
y = T[:,1]
fm = train(X, y)
X_new = sparse([
0 1 0 1 0 0 0 0 1 13.0;
])'
p = predict(fm, X_new)(X_train, y_train) = read_libsvm("data/small_train.libfm")
fm = train(sparse(X_train), y_train)
(X_test, y_test) = read_libsvm("data/small_test.libfm")
p = predict(fm, sparse(X_test))- AdaGrad SGD optimization
- Performance benchmark with libfm and python implementation (pyfm/fastFM/lightfm)
- Field-Aware FM
- Gaussian Process FM
- MCMC and ALS - just like in libfm