tensor-fm

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.

Install

The latest development version of tensorfm can be installed from its github repo with:

pip install git+https://github.com/gmodena/tensor-fm

Usage examples

Tensorlow and scikit-learn APIs are provided.

Tensorflow

The tensorflow implementation of Factorization Machines lives under tensor-fm/tensorfm/base.py. An example of how to work with this API can be found in tensor-fm/tests/test_base.py.

Scikit-learn estimator

tensorfm.sklearn exposes two sklearn compatible estimators: FactorizationMachineRegressor and FactorizationMachineClassifier.

Example

from tensorfm.sklearn import FactorizationMachineRegressor
...
fm = FactorizationMachineRegressor()
fm.fit(X, y)
fm.predict(X)

See also examples/movielens.py

Performance

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 from http://nowave.it/factorization-machines-with-tensorflow.html. 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 True.

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 of iterations.