This is a TensorFlow implementation of an arbitrary order (>=2) Factorization Machine based on paper Factorization Machines with libFM.

It supports:

• dense and sparse inputs
• different (gradient-based) optimization methods
• classification/regression via different loss functions (logistic and mse implemented)
• logging via TensorBoard

The inference time is linear with respect to the number of features.

Tested on Python3.5, but should work on Python2.7

This implementation is quite similar to the one described in Blondel's et al. paper [https://arxiv.org/abs/1607.07195], but was developed independently and prior to the first appearence of the paper.

Dependencies

• scikit-learn
• numpy
• tensorflow 1.0
• tqdm

Usage

The interface is similar to scikit-learn models. To train a 6-order FM model with rank=10 for 100 iterations with learning_rate=0.01 use the following sample

from tffm import TFFMClassifier
model = TFFMClassifier(
    order=6,
    rank=10,
    optimizer=tf.train.AdamOptimizer(learning_rate=0.01),
    n_epochs=100,
    batch_size=-1,
    init_std=0.001,
    input_type='dense'
)
model.fit(X_tr, y_tr, show_progress=True)

See example.ipynb and gpu_benchmark.ipynb for more details.

It's highly recommended to read GSOCFactorizationMachinesSystemML2/core.py for help.

Testing

Just run python test.py in the terminal. nosetests works too, but you must pass the --logging-level=WARNING flag to avoid printing insane amounts of TensorFlow logs to the screen.