Unlearning Protected User Attributes in Recommendations with Adversarial Training

This repository accompanies the paper Unlearning Protected User Attributes in Recommendations with Adversarial Training by Christian Ganhör, David Penz, Navid Rekabsaz, Oleg Lesota and Markus Schedl.

Below we give an overview of the most important aspects to reproduce our results.

Content

• Source code
• Running experiments
  • Preconditions
  • Training
  • Evaluation
  • TensorBoard

Supported datasets

Currently, this project focuses on two datasets

• LFM2b-bias
• MovieLens-1m

which both contain user-sensitive features such as gender, age and country information.

Note: Currently, only the gender attribute is supported in the experiments.

Both datasets need to be preprocessed, before they can be used in this project.
For preprocessing the LFM2b dataset, please check out the preprocessing as it was done in the datasets corresponding paper.

For the MovieLens-1m dataset, check out the notebook /notebooks/preprocess_movielens_1m.

After the datasets are prepared, adjust the specific paths in the global config file conf.py accordingly.

Additional datasets

Additional datasets can be used by

1. applying similar preprocessing (preferably as in /notebooks/preprocess_movielens_1m)
2. adding additional path variables in conf.py
3. adjusting the supported datasets in src/utils/input_validation.py
4. extend the 'if'-cases by the new dataset in src/utils/nn_utils/get_datasets_and_loaders()

Running experiments

Preconditions

• setup environment

  • conda env create -f bias_research.yml
    (bias_research_win.yml for Windows)

  • conda activate bias-research

  • python3 setup.py develop

Running experiments

Please check out the dedicated page on EXPERIMENTS
(You can also find the descriptions of the used training configurations there.)

Tensorboard

Many important aspects of training and validation are logged via TensorBoard. They can be viewed by opening TensorBoard in the results folder for a certain experiment.

These results folders are per default created in results/, in which each dataset has a separate folder. To change this path, please adjust the variable LOG_DIR in conf.py.

1. move into the experiments results folder, e.g.
   cd results/lfm2b/vae/standard--2021-07-15 14:14:25.366509

2. open TensorBoard
   tensorboard --logdir=./