A Counterfactual Approach for Addressing Individual User Unfairness in Collaborative Recommender System

A team of researchers has developed a novel method to tackle a core flaw in modern recommender systems: individual user unfairness. In a paper titled "A Counterfactual Approach for Addressing Individual User Unfairness in Collaborative Recommender System," Nikita Baidya, Bidyut Kr. Patra, and Ratnakar Dash address the problem where traditional collaborative filtering (CF) models provide poorer quality recommendations to certain users, a problem known as Individual User Unfairness Problem (IUUP). While prior research could identify this bias, it lacked effective solutions. The team's new approach directly mitigates the issue, which can lead to user dissatisfaction and business losses.

Their proposed solution is a dual-step, counterfactual methodology. First, it identifies candidate users who are receiving subpar recommendations. Then, it strategically introduces simulated, or 'counterfactual,' interactions for these users one at a time. By analyzing the benefit of this perturbation, the model learns to update its user and item embeddings more effectively, leading to fairer and more engaging recommendations for all. The researchers validated their approach on three standard datasets—MovieLens-100K, Amazon Beauty, and MovieLens-1M—demonstrating its superiority over existing techniques that only diagnose but don't fix the unfairness.