This journal paper comprehensively addresses the effect of popularity in false-positive metric measurements in the evaluation of recommender systems. In doing so, this work extends another work by the same authors, accepted for publication at SIGIR 2020
E. Mena-Maldonado, R. Cañamares, P. Castells, Y. Ren and M.Sanderson. 2021. Popularity Bias in False-Positive Metrics for Recommender Systems Evaluation. Transaction On Information Systems (TOIS).
Paper DOI (https://doi.org/10.1145/3452740) TOIS PAPER: TOIS2021 paper
This project contains two modules:
- Recommendation: we used (an edited version) of Librec 2.0.0 library to run the algorithms of our experiments (See librec-2.0.0 folder)
- Evaluation: we created some scripts in Python to carry out evaluation (See FP_metrics folder)
We have included instructions (README files) on how to run each module, please refer to each folder for more information.
For convinience, we have uploaded binarized versions of the datasets used for all the experiments presented in the paper. Please see the folder tois2021/librec-2.0.0/data
Dataset | Train | Test |
---|---|---|
MOVIELENS 1M | Observed | Observed |
CM100K | Observed | Observed and True |
CM100K SYNTHETIC | Observed | Observed and True |
YAHOO! R3 | Observed | Observed and True |
We run evaluation under 5-fold cross validation except in Movielens 1M time split.
The code was tested on Linux
NAME="Red Hat Enterprise Linux Server"
VERSION="7.7 (Maipo)"
Can possibly run on OSX however this has not been tested yet.