lrann

Experiments on the effectiveness of low-rank approximations in collaborative filtering compared to neural networks.

Installation

In order to reproduce our experiments:

create an environment lrann with the help of Miniconda,
```
conda env create -f environment.yaml
```
activate the new environment with
```
 conda activate lrann
```
install lrann with:
```
 python setup.py install # or `develop`
```
optionally run the unit tests by executing pytest

Then take a look into the experiments folder.

Run Experiments

In order to reproduce our research results, we provide an easy way to run different experiments on your own having the provided package installed. Each command requires three command line arguments:

-e: denotes the name of the experiment, see below
-c: config_file: relative path to the config file as already provided in experiments
-o: results_file: path where the results .csv-file should be saved

In addition, by adding -v you may enable verbose mode.

Best Neural Network Search

Run the comparative experiments between MF and DNN invoking the following command:

run_dnn_experiment -e nn_search -c <config_file> -o <results_file> -v

For example:

run_dnn_experiment -c experiments/experiment_config.yml -o test_result.csv -v

Matrix Factorization Hyperparameter Optimization

Run the matrix factorization hyperparameter search:

run_dnn_experiment -e mf_hyperopt -c <config_file> -o <results_file> -v

Covariance Analysis

In order to retrieve results for the covariance analysis, perform the following command

run_dnn_experiment -e covariance -c <config_file> -o <results_file> -v

Note

This project has been set up using PyScaffold 3.1. For details and usage information on PyScaffold see https://pyscaffold.org/.

The basic structure and some code was taken from the Spotlight recommender library, which is also MIT licensed.

Todo

Convert Numpy docstring style to Google style
Change command name from run_dnn_experiment to run_lrann_experiment

FlorianWilhelm/lrann