Scholarlink

<p>Graph Convolutional Neural Networks (GCNs) have opened new avenues for improving the quality of recommendations in collaborative filtering systems. Nevertheless, many existing models tend to amplify popularity bias when employing graph convolution methods to enhance recommendation accuracy. This often leads to unfair recommendations and inadequate representation of long-tail items. To tackle this problem, we introduce DPopGCF, an innovative graph convolutional collaborative filtering model aimed at reconciling accuracy with fairness. This model integrates two complementary innovative modules: (1) Flexible Negative Sampling (FNS), which dynamically adjusts the difficulty of selecting negative samples during training. This not only mitigates false positive and false negative instances but also reduces the training bias toward popular items by increasing the sampling probability of long-tail items; (2) Degree-Popularity-Aware Debiasing Mechanism (DPAD), which dynamically intervenes in the embedding updates of high-impact nodes during neighborhood aggregation. This mechanism mitigates the popularity bias amplified by graph convolution while ensuring that the exposure of popular items is not compromised, thereby increasing the recommendation frequency of long-tail items and balancing recommendation diversity and fairness. Experiments conducted on three publicly available datasets demonstrate that DPopGCF markedly enhances fairness without sacrificing accuracy, with average improvements of 1.96%, 1.57%, 2.35%, and 2.39% in NDCG, Recall, Precision, and F1 metrics, respectively, compared to leading baseline algorithms. This research provides a novel approach and perspective for fostering fairness in recommender systems.</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

Dpopgcf: a fairness-driven graph convolutional collaborative filtering model with flexible negative sampling

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Xiaoyang Wen,
  • Shaopeng Guan

摘要

Graph Convolutional Neural Networks (GCNs) have opened new avenues for improving the quality of recommendations in collaborative filtering systems. Nevertheless, many existing models tend to amplify popularity bias when employing graph convolution methods to enhance recommendation accuracy. This often leads to unfair recommendations and inadequate representation of long-tail items. To tackle this problem, we introduce DPopGCF, an innovative graph convolutional collaborative filtering model aimed at reconciling accuracy with fairness. This model integrates two complementary innovative modules: (1) Flexible Negative Sampling (FNS), which dynamically adjusts the difficulty of selecting negative samples during training. This not only mitigates false positive and false negative instances but also reduces the training bias toward popular items by increasing the sampling probability of long-tail items; (2) Degree-Popularity-Aware Debiasing Mechanism (DPAD), which dynamically intervenes in the embedding updates of high-impact nodes during neighborhood aggregation. This mechanism mitigates the popularity bias amplified by graph convolution while ensuring that the exposure of popular items is not compromised, thereby increasing the recommendation frequency of long-tail items and balancing recommendation diversity and fairness. Experiments conducted on three publicly available datasets demonstrate that DPopGCF markedly enhances fairness without sacrificing accuracy, with average improvements of 1.96%, 1.57%, 2.35%, and 2.39% in NDCG, Recall, Precision, and F1 metrics, respectively, compared to leading baseline algorithms. This research provides a novel approach and perspective for fostering fairness in recommender systems.