计算机科学与探索2025，Vol.19Issue(10)：2667-2682,16.DOI:10.3778/j.issn.1673-9418.2412024
基于分布感知优化的高鲁棒推荐方法
Distribution-Aware Optimization for Robust Recommendations
摘要
Abstract
With increasing applications of personalized recommendation systems on a variety of platforms,the need to ac-curately understand and model complex user behaviors along with vast item information has become critical.Traditional recommendation systems often overlook difficult samples that arise from user curiosity or misoperations.These challeng-ing samples,if not properly addressed,can lead to model bias and a significant drop in performance.Additionally,tradi-tional recommendation systems often only consider individual user-item interactions,failing to capture higher-order rela-tions from the perspective of distributions.To address the above limitations,this paper proposes a novel distribution-aware optimization for robust recommendations(DORRec),which targets the matching between global user and item dis-tributions without supervision while distinguishing difficult samples to achieve robust recommendations.Specifically,in the distribution-based difficult sample distinguishing module,this paper relaxes the regularization constraints under the Sinkhorn distance to compute closed-form solutions for complex user-item matching scores,so as to find the hard samples for each user.Furthermore,in the adaptive threshold-based high-robustness recommendation module,this paper proposes a personalized threshold mechanism that adaptively adjusts interaction weights to enhance the training of difficult sam-ples,meeting the need for robust recommendations.Experiments on four public datasets validate the effectiveness of the DORRec framework,demonstrating significant improvements in accuracy and robustness.Compared with multiple state-of-the-art recommendation algorithms and components on several evaluation metrics,DORRec demonstrates significant superiority in recommendation performance.关键词
推荐系统/困难样本/分布感知/鲁棒推荐/最优传输Key words
recommendation system/difficult samples/distribution-aware/robust recommendations/optimal transport分类
信息技术与安全科学引用本文复制引用
檀彦超,周子皓,马国芳,王石平,黄维,阳及,李天瑞..基于分布感知优化的高鲁棒推荐方法[J].计算机科学与探索,2025,19(10):2667-2682,16.基金项目
国家自然科学基金(62302098) （62302098）
浙江省自然科学基金(LQ23F020007) （LQ23F020007）
浙江省农业农村厅项目(2024SNJF044) （2024SNJF044）
福建省自然科学基金面上项目(2025J01540).This work was supported by the National Natural Science Foundation of China(62302098),the Natural Science Foundation of Zhejiang Province(LQ23F020007),the Project of Zhejiang Provincial Department of Agriculture and Rural Affairs(2024SNJF044),and the Nat-ural Science Foundation of Fujian Province(2025J01540). （2025J01540）
