Beyond TF-IDF: Reproducibility and Generalizability of Two-sided Fairness Tradeoffs
Abstract: In this paper, we reproduce the experiments conducted by Greenwood et al. (2024) to validate their findings. We successfully recreate their main empirical results and conclude that the authors' two primary claims are valid: (a) item fairness tends to impose a higher cost on user fairness in more homogeneous populations, and (b) the cost of misestimation is high and item fairness constraints do not affect this cost. We extend their experiments by using SPECTER embeddings and the Amazon Book dataset to validate the generalizability of their results. We find similar results, for both the original dataset and the Amazon Book dataset, consistent with those of the authors. However, the results obtained with SPECTER embeddings do not fully support the two claims.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Hongfu_Liu2
Submission Number: 4285
Loading