Genetic Fairness Optimization in Neural Fair Recommenders: Beyond Accuracy and Diversity

Abstract

A Recommender System is a kind of complete and full-featured pipeline that specializes in personalizing online experiences related to e-commerce, media, and educational applications. Popular online platforms such as YouTube, Amazon, and Netflix utilize a recommendation system that suggests items and products to the consumer depending on their past behavior and history. The accuracy of the recommended items and services is often undermined due to the biases of the data. Moreover, recommendations often have lower diversity. This thesis proposes a multi-stage fairness-aware recommendation pipeline where Neural Collaborative Filtering (NCF) is used as the baseline model. After that, embedding debiasing, fairness-aware fine-tuning are implemented, and lastly, Genetic Algorithm (GA)-based re-ranking is used as the last layer to compare accuracy, fairness gap, and diversity to find out the best result. The approach is evaluated on large-scale MovieLens datasets (10M, 20M, and 32M) using ranking metrics, fairness measures, and diversity indicators. Results show that embedding debiasing and fairness-aware fine-tuning reduces disparities across user groups, while GA-based re-ranking improves item exposure equity and catalog diversity with minimal loss in accuracy. The findings demonstrate that fairness, diversity, and accuracy can be jointly optimized, and that evolutionary genetic optimization provides a scalable mechanism to achieve Pareto-optimal trade-offs. This proposed pipeline also demonstrates how accuracy, fairness gap, and diversity change as the dataset is gradually increased. This contribution of work provides a useful framework to construct a recommendation system that not only performs well but also remains trustworthy and unbiased.