FaiRIR : A suite of Fair Related Item Recommendation Alogorithms

Code base for the paper titled, FaiRIR: Mitigating Exposure Bias from Related Item Recommendations in Two-Sided Platforms to appear at IEEE Transactions on Computational Social Systems 2022 (IEEE TCSS'22).

If you are using this code for any research publication, please cite the following paper as the source of the code.

Abhisek Dash, Abhijnan Chakraborty, Saptarshi Ghosh, Animesh Mukherjee and Krishna P. Gummadi. "FaiRIR: Mitigating Exposure Bias from Related Item Recommendations in Two-Sided Platforms”. To appear at IEEE Transactions on Computational Social Systems 2022 (IEEE TCSS'22).

BibTex

@article{dash2022fairir,
title={FaiRIR: Mitigating Exposure Bias from Related Item Recommendations in Two-Sided Platforms},
author={Dash, Abhisek and Chakraborty, Abhijnan and Ghosh, Saptarshi and Mukhejree, Animesh and Gummadi, Krishna P.},
journal={IEEE Transactions on Computational Social Systems},
year={2022},
publisher={IEEE}
}

URL to the datasets:

• MovieLens: https://grouplens.org/datasets/movielens/
• Amazon Review Dataset: http://jmcauley.ucsd.edu/data/amazon/links.html

Preprocessing files:

• ratingExctracter.py shall extract the ratings from the data files obtained from the above URLs.

Vanilla representation learning:

• VanillaSVDRepresentation.py (VanillaI2VRepresentation.py) needs to be run to learn the vanilla representation of items / movies for further processing.
• input: user-item interaction logs
• output: item representations

Similarity evaluation:

• simeval.py (similarity_item2vec.py) evaluates the dense similarity between all pairs of item in the datasets.
• input: learnt representations
• output: A dense matrix of similarity between items

Related Item Recommendation generation:

• RIR.py generates the recommendation instances provided the similarity matrix obtained in similarity evaluation.
• input: similarity matrix
• output: Related Item Recommendations for each item

Related Item Network generation:

• RIN.py generates the related item recommendation as discussed in the paper from the related item recommendations. It also performs the exposure analyses of the RINs and provides the exposure bias score.
• input: related item recommendations
• output: related item network, exposure bias score

FairRIR neighbor algorithm:

• FaiRIR_nbr.py generates the fair recommendation network from the given desired exposure distribution and relatedness of items.
• input: similarity matrix among items, desired exposure of items
• output: Fair related item recommendation network

FairRIR representation learning:

• ./FaiRIR_RL/main.py (main_i2v.py) generates the fair representations learnt by optimizing for relatedness and desiredness loss. For more details, please refer to the paper.
• input: vanilla representations of algorithms, desiredness graph representations
• output: fair representations
• To generate FaiRIR recommendations, please use the RIR.py with new input being the fair representations learnt above.

Please send any questions you might have about the code and/or the algorithm to [email protected].