bibs2091/tfmf

Collaborative Filtering Using Matrix Factorization in TensorFlow

This package implements MatrixFactorizer class for collaborative filtering using matrix factorization with implicit (Hu et al, 2008) and explicit ratings (Koren et al, 2009, Yu et al, 2012). The MatrixFactorizer class works with sparse data (scipy.sparse matrices) and uses TensorFlow as a workhorse. The trained models can be easily saved and restored.

Example

Below an example of using MatrixFactorizer class for explicit ratings is shown. The data is converted to to scipy.sparse.dok_matrix format using the sparse_matrix function provided in the package as a convenience wraper.

import numpy as np
from tfmf import MatrixFactorizer, sparse_matrix

user_id = [0,0,1,1,2,2]
movie_id = [0,1,2,0,1,2]
rating = [1,1,2,2,3,3]
X = sparse_matrix(user_id, movie_id, rating)

The model is trained using the fit method. We choose to use two latent dimmensions and random batches of size 4 for the training. To make predictions from the model, we use predict_all method that makes predictions for the whole factorized matrix (or slices of it), but more standard choice would be to use predict(user_ids, item_ids) method for making predictions for individual users and items.

mf = MatrixFactorizer(n_components=2, batch_size=4, implicit=False)
mf.fit(X)
mf.predict_all().A
## array([[1.013387 , 1.0056534, 1.11131  ],
##        [1.926976 , 1.8363876, 1.9962051],
##        [1.9263643, 2.969007 , 2.9673567]], dtype=float32)
X.A
## array([[1, 1, 0],
##        [2, 0, 2],
##        [0, 3, 3]])

The learned parameters are saved to temporary directory.

import tempfile
tmpdir = tempfile.gettempdir()
mf.save(tmpdir + '/tfmf')

Next, we initialize the class once again. Using the init_with_shape method, MatrixFactorizer class is set-up to model the matrix of shape (n_users, n_items) with randomly initialized parameters. Predictions from such model are meaningless, because it wasn't trained yet.

mf = MatrixFactorizer(n_components=2, batch_size=4, implicit=False)
mf.init_with_shape(3, 3)
mf.predict_all().A
## array([[ 4.6303128e-05, -5.7351419e-05,  1.7567796e-05],
##        [-1.0485943e-04,  6.2389910e-05, -3.7540267e-05],
##        [-1.2580390e-04, -2.4326339e-04, -3.4460012e-05]], dtype=float32)

Instead of training, we resore the previously learned parameters. Such restored model can be trained further using the partial_fit method, or used to make predictions.

mf.restore(tmpdir + '/tfmf')
mf.predict_all().A
## array([[1.013387 , 1.0056534, 1.11131  ],
##        [1.926976 , 1.8363876, 1.9962051],
##        [1.9263643, 2.969007 , 2.9673567]], dtype=float32)

References

Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8).

Yu, H. F., Hsieh, C. J., Si, S., & Dhillon, I. (2012, December). Scalable coordinate descent approaches to parallel matrix factorization for recommender systems. In Data Mining (ICDM), 2012 IEEE 12th International Conference on (pp. 765-774). IEEE.

Hu, Y., Koren, Y., & Volinsky, C. (2008, December). Collaborative filtering for implicit feedback datasets. In Data Mining, 2008. ICDM'08. Eighth IEEE International Conference on (pp. 263-272). IEEE.