mmosc / recommendation_systems_fairness

This repository accompaines the paper "Investigating Gender Fairness of Recommendation Algorithms in the Music Domain"

Investiating Gender Fairness of Recommendation Algorithms in the Music Domain

This repository accompanies the paper "Investigating Gender Fairness of Recommendation Algorithms in the Music Domain" by Alessandro B. Melchiorre, Navid Rekabsaz, Emilia Parada-Cabaleiro, Stefan Brandl, Oleg Lesota, and Markus Schedl.

Installation

Create the environment with

conda env create -f gender_fair.yml
conda activate gender_fair

and then setup the project with

python3 setup.py develop

Setup the data

Data is available at http://www.cp.jku.at/datasets/LFM-2b/

To sample the items as done in the paper, please check item_sampler.ipynb in notebooks folder.

Data paths must be set in the conf.py file. In particular:

• DATA_PATH: points to a .tsv file with columns [user_id,track_id,play_count].
• DEMO_PATH: points to a .tsv file with columns [user_name,country,age,gender,timestamp].
• TRACKS_PATH: points to a .tsv file with columns [track_artist,track_name].
• DOWN_DATA_PATH: similar to DATA_PATH but contains the balanced dataset (optional).
• DOWN_DEMO_PATH: similar to DEMO_PATH but contains the balanced dataset (optional).
• OUT_DIR: where the processed datasets (e.g train/val/test data for fold 0).

Run an experiment (Train + Validation)

algorithms contains all the algorithms used in the project. Move to the folder of the algorithm of your choice (e.g. slim)

cd algorithms/slim

and with the active env, run the experiment

python3 main_slim.py --experiment_type <experiment_type>

experiment_type can be:

• standard: original data
• up_sample: debiasing method where the minority group in the training data is upsampled (as specified in the paper)
• down_sample: balanced data is used instead where users from the majority groups have been discarded

for vae there is also the parameter gpu which specifies which gpu to use.

Check the experiment status

Experiment status and results will be saved in the res folder (created when the first experiment is run).

To check the status of your experiment, move in the folder of the algorithm you are running (e.g. slim):

cd res/slim/<experiment_type>-<timestamp_start_experiment>/val

For a broad overview of the experiments, you can just use tensorboard in the folder.

tensorboard --logdir=./

You can also jump to a specific test fold by selecting one of the folders there (numbered from 0 to 4 for 5-folds cross validation). E.g. monitor how the experiment is going on fold 0.

cd 0
tensorboard --logdir=./

Running evaluation

Experiment status and results will be saved in the res folder (created when the first experiment is run).

To run the evaluation when an experiment has ended, move in the folder of the algorithm you run (e.g. slim)

cd algorithms/slim

and the run the evaluation with:

python3 eval_slim.py --experiment_type <experiment_type>

where experiment_type parameter is defined as above in "Check the experiment status".

NOTE: Evaluation requires loading the trained model (except for pop). The code will look in the res folder for the best configuration automatically. It is needed to update the path to the trained model by providing the timestamp_start_experiment in the evaluation code. E.g. for eval_slim.py you can find this at the top of the code:

best_configs = {
    'standard': '2020-11-24 19:43:47.801230',
    'up_sample': '2020-11-24 19:44:32.188924',
    'down_sample': '2020-11-24 19:45:15.744244'
}

Set the timestamp of your experiment in the dictionary (in the correct experiment_type field) for running the evaluation.

E.G. slim is started at `2020-11-24 19:43:47.801230 with

python3 main_slim.py --experiment_type standard

in res there will be a folder slim/standard-2020-11-24 19:43:47.801230. After the training and validation, place 2020-11-24 19:43:47.801230 in the filed standard of the best_configs dictionary in eval_slim.py

Results and best configurations

Experiment status and results will be saved in the res folder (created when the first experiment is run).

Results

To check the results of your experiment, move in the folder of your experiment (e.g. slim):

cd res/slim/<experiment_type>-<timestamp_start_experiment>/test

For a broad overview of the experiments, you can just use tensorboard in the folder.

tensorboard --logdir=./

The results are saved separately for each fold in the relative folder. In each fold folder you will find:

full_metrics.pkl
full_raw_metrics.pkl

Both files contain a dictionary where the key is the evaluated metric function and the value is the result of the metric function on the test data.

• full_raw_metrics.pkl contains vectors as values of the dictionary (results are not aggregated).
• full_metrics.pkl contains floats as values of the dictionary (results are aggregated with mean)

Best configurations

The best configuration for each algorithm and for each fold is located in:

cd res/slim/<experiment_type>-<timestamp_start_experiment>/val/<fold-number>/best_config.pkl

Significance Testing

The significant testing is located in signigicance_testing folder. Code is in R and Jupyter notebook. For instructions, refer to the README in the folder.

cd significance_testing