In this repository, we have uploaded the code corresponding to the SIGIR2020 Paper "Controlling Fairness and Bias in Dynamic Learning-to-Rank", which is also on arxiv (https://arxiv.org/abs/2005.14713).
The Paper contains 10 Figures corresponding to 10 different experiments. You can reproduce all of them from the command line.
To start Experiment 1, call
python main.py with 'EXPERIMENT=1' 'PLOT_PREFIX="plots/Exp1/"'
Experients 1-6 are based on a Real News Dataset. This can be downloaded as a csv file from adfontes media and must be saved in as "InteractiveMediaBiasChart.csv" in the data subfolder.
Experients 7-10 are based on the MovieLens 20M Dataset which must be saved in as "ratings.csv" in the data subfolder.
data_utils.py: For loading and preprocessing the data
Experiments.py: Here we define the functions called for different experiments (like different starting users, different user populations, etc.)
Documents.py: Just the classes for our Items (Movie, News article)
relevance_network.py: The neural network for the personalized ranking
Simulation.py : Here is the simulation happening, incl. Ranking Function, Click_Model, ...
To get started, take a look at the Jupyter Notebook "Example.ipynb". Here you can run a single trial for all our Ranking-Methods on both the News and the Movie Dataset