Factorization machine in TensorFlow 2

Example implementing a factorization machine in TensorFlow 2, along with a framework for generating user-item ratings for testing.

Based on the theory and tf-1 code in this article by Gabriele Modena and the paper by Factorization Machines by Steffen Rendle. Please refer to these articles for much more detailed descriptions of the model, and the maths that avoid O(p²).

Code is work in progress so expect bugs and inconveniences. See examples/ directory for more detailed examples.

Install

git clone https://github.com/garethjns/TF2FactorizationMachine.git

Example

TensorFlow 2 model

For now the current model only implements .__init__ .__call__ and follows a similar structure as the TensorFlow linear models example. Training is done in a loop with a function (train_step) to update the parameters; these will be moved to .fit at some point.

See also examples/1_model_development.ipynb for mode details on model development and examples/2_movie_lens_data.ipynb for an example running it on the MovieLens dataset.

import numpy as np

import matplotlib.pyplot as plt

from fmachine.model import FactorizationMachine
from fmachine.helpers import train_step, l2_loss

# Features
x = np.array([[1, 0, 0,  1, 0, 0, 0,  0.3, 0.3, 0.3, 0.0,  13,  0, 0, 0, 0 ],
              [1, 0, 0,  0, 1, 0, 0,  0.3, 0.3, 0.3, 0.0,  14,  1, 0, 0, 0 ],
              [1, 0, 0,  0, 0, 1, 0,  0.3, 0.3, 0.3, 0.0,  16,  0, 1, 0, 0 ],
              [0, 1, 0,  0, 0, 1, 0,  0.0, 0.0, 0.5, 0.5,  5,   0, 0, 0, 0 ],
              [0, 1, 0,  0, 0, 0, 1,  0.0, 0.0, 0.5, 0.5,  8,   0, 0, 1, 0 ],
              [0, 0, 1,  1, 0, 0, 0,  0.5, 0.0, 0.5, 0.0,  9,   0, 0, 0, 0 ],
              [0, 0, 1,  0, 0, 1, 0,  0.5, 0.0, 0.5, 0.0,  12,  1, 0, 0, 0 ]])

# Targets (explicit rating)
y = np.array([5, 3, 1, 4, 5, 1, 5])
y.shape = (7, 1)

mod = FactorizationMachine(m=16)

# Training
epochs = 200
bs, ws, vs, losses = [], [], [], []
for e in range(epochs):
    cur_loss = train_step(mod=mod, 
                     x=x, 
                     y_true=y,
                     lr=0.0025,
                     loss_f=l2_loss)

# PLot loss
plt.plot(losses)
plt.title('Loss history')
plt.xlabel('Epoch')
plt.ylabel('Loss')

Keras interface

Todo

Data generator

Todo