Probabilistic Matrix Factorization

This is an implementation of PMF algorithm.

For the idea of this algorithm, please see Probabilistic Matrix Factorization, Ruslan Salakhutdinov and Andriy Mnih.

If you meet any problem while running the code, please file an issue or contact me by loginaway@gmail.com.

Dependencies

Only numpy is needed to be installed manually. Try

pip install numpy

or by conda, try

conda install numpy

Usage

Example Usage

1, modify the configuration in pmf.conf. For the details, see Configuration.

2, run python -m PMF in the parent directory of PMF.

Dataset

Your dataset should be formulated into a .txt file, with each row being an item of data like 0, 1, 5 , where 0 is the index of a certain user, and 1 is that of a movie, and 5 the rating user-0 gives on movie-1.

Configuration

You are required to modify pmf.conf if you want to personalize the settings. There are some explanations for the options.

trainset_name

testset_name

​ The training dataset name and the testing dataset name.

stepsize

​ The step size used in Stochastic Gradient Descent method.

embedding_dimension

​ The column size of matrix U and V. (The row sizes are the number of users and number of movies, respectively, set by UVsize.)

epoch_num

​ The maximal number of iteration you want the optimizing algorithm to run.

k_u

k_v

​ The weights of the regularization.

UVsize

​ The row size of matrix U and V, it is a tuple of dimension 2. For example

​ UVsize=(5, 6) will assign 5 rows to U and 6 rows to V.

ndcg_top

​ Return ndcg evaluation results. For example

​ ndcg_top=1, 3, 5 will generate NDCG@1, 3, 5.

rmse

​ Whether or not to generate RMSE evaluation. rmse=True / rmse=False

References

[1] Ruslan Salakhutdinov and Andriy Mnih, Probabilistic Matrix Factorization, 2007.

​