融合迁移学习和解纠缠负采样的去偏推荐方法  

Debiasing recommendation method integrating transfer learning and disentangled negative sampling

作　　者：刘琦 唐宏[1,2] 杨力鸣 Liu Qi;Tang Hong;Yang Liming(School of Communications&Information Engineering,Chongqing University of Posts&Telecommunications,Chongqing 400065,China;Key Laboratory of Mobile Communications Technology of Chongqing,Chongqing University of Posts&Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学移动通信技术重庆市重点实验室,重庆400065

出　　处：《计算机应用研究》2025年第4期1115-1121,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(61971080)。

摘　　要：针对现有去偏推荐方法在选择负样本时将样本作为一个整体考虑导致的采样偏差问题,以及不平衡的热门-长尾项目表征学习无法有效缓解数据稀疏的问题,提出融合迁移学习和解纠缠负采样的去偏推荐方法(DTDN)。该方法首先利用交互行为中的对撞效应设计负采样模块;其次,根据采样数据设计特征解耦模块对用户和正负样本的特征进行解耦表征学习(DRL);然后,在表征学习阶段引入迁移学习模块,以对齐热门项目和长尾项目的表征分布;最后,基于解耦表征设计样本选择器去除数据中的固有偏差,帮助模型准确学习用户和正负样本特征中的规律。在两个实际应用数据集上的广泛实验结果表明,与最优的基线方法CD 2AN相比,DTDN的各项性能指标均有明显的提升,验证了DTDN在降低推荐系统偏差负面影响方面的有效性。To address the sampling bias caused by treating negative samples holistically in existing debiased recommendation methods,along with the limitations of imbalanced representation learning for popular and long-tail items in alleviating data sparsity,this study proposed a debiased recommendation method integrating transfer learning and disentangled negative sampling(DTDN).The proposed method firstly designed a negative sampling module based on the collision effect observed in user interactions.Then,it developed a feature disentanglement module to perform disentangled representation learning(DRL)on user features and both positive and negative samples.During representation learning,a transfer learning module aligned the distributions of popular and long-tail items,mitigating sparsity issues.Finally,a sample selector based on the disentangled representations was constructed to remove inherent data biases,enabling the model to effectively capture patterns in user pre-ferences and sample features.Experimental results on two real-world datasets demonstrate that DTDN significantly improves all performance metrics compared to the best baseline,CD 2AN,verifying its effectiveness in reducing the negative impact of bias in recommendation systems.

关 键 词：推荐系统 去偏推荐 负采样 解纠缠表征学习 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]