| 摘 要: 现有的去偏方法多采用全局同质化的惩罚策略,忽略了用户对流行度偏好存在的异质性。这种全局干预导致在提升系统公平性的同时易造成推荐准确率的显著下降,难以在准确性与公平性之间获取更优的帕累托权衡(Pareto Trade-off)。为此,本文在现有假阳性纠错机制的基础上,提出了一种用户自适应图结构去偏扩展方法(User-Adaptive Graph-FPC,简称UAG-FPC)。该方法以轻量级图卷积网络(LightGCN)为骨干模型,首先量化用户表观从众度,并将其非线性映射为个性化的去偏惩罚系数；结合物品的实时度数构建基于物品度数的动态边权重衰减策略。本文在贝叶斯最大后验概率估计(Maximum A Posteriori Estimation,MAP)框架下,推导该动态惩罚项与一种从众先验的 MAP 估计等价性。基于两个公开数据集的长周期动态仿真实验表明,User-Adaptive Graph-FPC 成功克服了传统全局干预的局限性,实验拓展了去偏推荐的帕累托前沿。该方法在有效降低基尼因子和促进推荐公平性的同时,充分保留了用户的真实点击收益,为准确性与去偏效果的权衡提供了更优的解决方案。 |
| 关键词: 推荐系统 流行性偏差 图卷积网络 自适应去偏 |
|
中图分类号:
文献标识码:
|
|
| Research on Debiased Recommendation Algorithms with User-Adaptive Graph Error Correction |
|
WANGSEN, ZOU HAITAO
|
Jiangsu University of Science and Technology
|
| Abstract: Most existing debiasing methods adopt globally homogeneous penalty strategies, ignoring the heterogeneity in users" popularity preferences. Such global interventions tend to cause excessive degradation in recommendation accuracy while improving system fairness, making it difficult to achieve a better Pareto trade-off between accuracy and fairness. To this end, based on the existing False Positive Correction (FPC) mechanism, this paper proposes a user-adaptive graph structure debiasing extension method, namely User-Adaptive Graph-FPC. Taking the Lightweight Graph Convolutional Network (LightGCN) as the backbone model, this method first quantifies users" historical herd tendency and nonlinearly maps it into personalized debiasing penalty coefficients. Combined with the real-time degrees of items, it further constructs a dynamic edge weight decay strategy based on item degrees. Within the framework of Bayesian Maximum A Posteriori (MAP) estimation, this paper discusses the theoretical motivation of this dynamic penalty term as an approximate solution incorporating a specific heuristic herd prior. Long-term dynamic simulation experiments on two public datasets demonstrate that User-Adaptive Graph-FPC successfully overcomes the limitations of traditional global interventions and extends the Pareto frontier of debiased recommendations. While effectively reducing the Gini coefficient and promoting recommendation fairness, this method sufficiently preserves users" real-click utilities, providing a superior solution for balancing accuracy and debiasing performance. |
| Keywords: Recommender Systems Popularity Bias Graph Convolutional Networks Adaptive Debiasing |
