This repository contains the code used for the experiments in "The Bandwagon Effect: Not Just Another Bias" published at ICTIR 2022 (preprint available).
If you use this code to produce results for your scientific publication, or if you share a copy or fork, please refer to our ICTIR 2022 paper:
@inproceedings{knyazev2022bandwagon,
Author = {Knyazev, Norman and Oosterhuis, Harrie},
Booktitle = {Proceedings of the 2022 ACM SIGIR International Conference on the Theory of Information Retrieval (ICTIR '22)}
Organization = {ACM},
Title = {The Bandwagon Effect: Not Just Another Bias},
Year = {2022}
}
The contents of this repository are licensed under the MIT license. If you modify its contents in any way, please link back to this repository.
This code makes use of Python 3 and the following packages: jupyter, matplotlib, numpy, scipy and tqdm. Make sure they are installed.
There are three files, which can be accessed by running jupyter notebook .
in the project folder.
lambda_estimation.ipynb
is used to obtain the (mis)estimated lambda values under a strong bandwagon effect, as described in Section 6.1.
figure3.ipynb
is used to generate Figure 3.
figure4.ipynb
is used to generate individual subfigures from Figure 4. Variables a
and b
can be used to modify the strength of the bandwagon effect. Variables a_hat
and b_hat
can be used to provide (mis)estimated lambda values to all estimators. In case of memory limitations, there is an option to calculate and save, or load predictions for a subset of estimators by running the cells marked as OPTIONAL
.