evfro / hybridsvd

Supplementary materials to reproduce experiments

Guide to reproduce experiments

If you just want to see all the results, simply open View_all_results.ipynb notebook. You can do it directly on Github. All other notebooks (with code) can also be viewed with Github.

Update for reviewers: The results for cold-start experiments are updated. Previously submitted results for the SVD models were based on suboptimal rank choice and, therefore, were lower than they should be. Confidence intervals are also included for all results in View_all_results.ipynb. The updated version of text can be found here.

Preparing environment for experimentation

NOTE: Only Linux environment is supported. The code has been tested on Ubuntu 14.04.

In order to run the code in this repository you firstly need to install the following software:

1. Python 2.7 (Python 3 is not supported).
   1.1 Pandas, Numpy, Scipy, Matplotlib, Jupyter and Numba packages (see instructions below).
2. Suitesparse (provides CHOLMOD functionality to compute sparse Cholesky decomposition).
3. Scikit-sparse package which conveniently wraps suitesparse (provided in the repository).
4. Special version of Polara framework, used to conduct all experiments (provided in the repository).
5. GraphLab Create to run Factorization Machines. This step can be skipped if you only want to test SVD-based models.

Python

The easiest (and recommended) way to get python and all required packages at once is to use the latest Anaconda distribution. If you use a separate conda environment for testing, the following command can be used to ensure that all required dependencies are in place (see this for more info):

conda install --file conda_req.txt

Alternatively, a new conda environment with all required packages can be created by:

conda create -n <your_environment_name> python=2.7 --file conda_req.txt

The file conda_req.txt can be found in polara_fixed.zip archive in this repository.

Suitesparse

Suitesparse can be installed with the following command:

sudo apt-get install libsuitesparse-dev

Scikit-sparse

A fixed version of scikit-sparse is provided within the repository and can be installed with pip (don't forget to activate your conda environment if you use it):

pip install --user scikit-sparse.zip

Polara

This is the most important part as it provides tools to conduct full experiment. Polara can also be installed with pip (mind your conda environment):

pip install --user --upgrade polara_fixed.zip

GraphLab Create

Simply follow the instructions on the download page https://turi.com/download/install-graphlab-create.html. You'll have to request an academic license to get a free version of this product.

Running the code

Use Jupyter Notebooks to play with the code. There are 4 main notebooks, corresponding to 4 experiments: standard and cold-start scenarios for Movielens (ML) and BookCrossing(BX) datasets. The names of the notebooks are self-explaining.