A scikit-learn compatible order 2 Factorization Machine, implemented atop TensorFlow 2. The algorithm is described in http://www.csie.ntu.edu.tw/~b97053/paper/Rendle2010FM.pdf. For an higher level overview of the method see http://nowave.it/factorization-machines-with-tensorflow.html.
This package is a port to Tensorflow 2 of the code presented in that blog post. The goal of this project is to experiment with different optimization strategies for classical ML models, and scalability of TF2 backends.
The latest development version of
tensorfm can be installed from its
github repo with:
pip install git+https://github.com/gmodena/tensor-fm
Tensorlow and scikit-learn APIs are provided.
The tensorflow implementation of Factorization Machines lives under
An example of how to work with this API can be found in
tensorfm.sklearn exposes two sklearn compatible estimators:
from tensorfm.sklearn import FactorizationMachineRegressor ... fm = FactorizationMachineRegressor() fm.fit(X, y) fm.predict(X)
All parameters and settings being equal, I noticed a considerable performance degradation of
FactorizationMachineRegressor (MSE on train/test) on movielens compared to the tensorflow 1 implementation
Possibly related, a test in the
check_regressors_train suite (
sklearn) fails due to a low
R^2. As a workaround
FactorizationMachineRegressor sets the
poor_score tag to
Limitations and known issues
Operations on sparse matrices are currently not supported.
Training continues till
max_iter is reached, we should stop if performance does not improve for a certain number