Scholarlink

<p>As fairness becomes an increasingly critical concern in recommender systems, evaluating models that balance accuracy and fairness remains a significant research challenge. This paper presents a comprehensive benchmarking study of fairness-aware recommender systems, ranging from classical matrix factorization techniques to state-of-the-art graph-based and neural models. Using the MovieLens 100&#xa0;K and 1&#xa0;M datasets, we evaluate eight models—MF, FairMF, FairRec, BiasExplain, NeuMF, MACR, LightGCN, and CERec—across accuracy metrics (AUC, F1, Precision, Recall, Test Loss) and fairness metrics (Statistical Parity, Group AUC, Equal Opportunity, Equalized Odds). Experimental results indicate that NeuMF achieves the highest AUC (0.8359 ± 0.0101), with CERec (0.8330 ± 0.1493) and LightGCN (0.8320 ± 0.0042) close behind; CERec and LightGCN offer competitive trade-offs between accuracy and fairness. Notably, LightGCN exhibits the most stable performance across seeds, whereas CERec provides a strong performance–fairness trade-off despite its higher variance. Among fairness-aware models, FairRec achieves the lowest Statistical Parity (0.0077 ± 0.0066), while BiasExplain and FairMF yield marginal fairness gains at the cost of accuracy. Visualization analyses (ROC, PR curves, PCA, and t-SNE) further reveal representational distinctions and clustering behaviors among models. Our findings highlight that high-performing models can still uphold fairness, and traditional fairness-aware approaches may require refinement. We conclude with recommendations for model deployment and future research in fairness-driven recommendation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking fairness-aware recommender systems from matrix factorization to graph-based models

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Hung Nguyen

摘要

As fairness becomes an increasingly critical concern in recommender systems, evaluating models that balance accuracy and fairness remains a significant research challenge. This paper presents a comprehensive benchmarking study of fairness-aware recommender systems, ranging from classical matrix factorization techniques to state-of-the-art graph-based and neural models. Using the MovieLens 100 K and 1 M datasets, we evaluate eight models—MF, FairMF, FairRec, BiasExplain, NeuMF, MACR, LightGCN, and CERec—across accuracy metrics (AUC, F1, Precision, Recall, Test Loss) and fairness metrics (Statistical Parity, Group AUC, Equal Opportunity, Equalized Odds). Experimental results indicate that NeuMF achieves the highest AUC (0.8359 ± 0.0101), with CERec (0.8330 ± 0.1493) and LightGCN (0.8320 ± 0.0042) close behind; CERec and LightGCN offer competitive trade-offs between accuracy and fairness. Notably, LightGCN exhibits the most stable performance across seeds, whereas CERec provides a strong performance–fairness trade-off despite its higher variance. Among fairness-aware models, FairRec achieves the lowest Statistical Parity (0.0077 ± 0.0066), while BiasExplain and FairMF yield marginal fairness gains at the cost of accuracy. Visualization analyses (ROC, PR curves, PCA, and t-SNE) further reveal representational distinctions and clustering behaviors among models. Our findings highlight that high-performing models can still uphold fairness, and traditional fairness-aware approaches may require refinement. We conclude with recommendations for model deployment and future research in fairness-driven recommendation.