A Probabilistic Greedy Attempt to Be Fair in Neural Team Recommendation

ABSTRACT Neural team recommendation has brought state‐of‐the‐art efficacy while enhancing efficiency at forming teams of experts whose success in completing complex tasks is almost surely guaranteed. However, they overlook fairness, that is, predicted teams are heavily biased toward popular and male experts, falling short of recommending female or non popular experts. In this work, we introduce and formalize the fair team recommendation problem in view of group‐based notions of fairness. Inspired by the promising performance of probabilistic rerankers in user‐item recommender systems for fairness guarantees, we further develop a probabilistic greedy reranking algorithm to achieve fairness with respect to popularity or gender biases in neural models based on different notions of fairness, including demographic parity and equal opportunity . Specifically, we aim to ensure a minimum representation of experts from the disadvantaged nonpopular or female groups by reranking the neural model's ranked list of recommended experts. Our experiments on three large‐scale benchmark datasets demonstrate: (1) neural team recommenders heavily suffer from biases toward popular and male experts; (2) our reranking method can substantially mitigate such biases while maintaining teams' efficacy; (3) in the presence of extreme biases in specific domains like gender disparities in US patents, post‐processing reranking methods alone fall short to demonstrate consistent mitigation performance across all fairness evaluation metrics, urging further tandem integration of pre‐process and in‐process debiasing techniques. The code to reproduce the experiments reported in this paper is available at https://github.com/fani‐lab/Adila/tree/coin25 .