基于融合奖励的神经协同过滤去曝光偏差推荐模型  

Neural collaborative filtering recommendation model forde-exposure bias based on fused rewards

作　　者：李鹏 李晓珊 朱心如 Li Peng;Li Xiaoshan;Zhu Xinru(School of Management,Harbin University of Commerce,Harbin 150028,China)

机构地区：[1]哈尔滨商业大学管理学院,哈尔滨150028

出　　处：《计算机应用研究》2025年第1期78-85,共8页Application Research of Computers

基　　金：2023年哈尔滨商业大学青年科研创新人才培育计划资助项目(2023-KYYWF-1001);黑龙江省博士后科研启动金资助项目(BS0053)。

摘　　要：推荐系统中因交互数据稀疏性和曝光不均导致的强曝光偏差,会集中推荐高曝光物品,忽略低曝光物品的潜在价值,从而限制用户选择并降低用户体验。为解决这一问题,提出一种结合神经协同过滤和线性置信上界算法的去曝光偏差模型。首先,通过分析用户与物品之间的交互数据,利用神经协同过滤算法学习用户和物品的特征,捕捉其潜在偏好;其次,引入线性置信上界算法,并将其生成的奖励值特征嵌入到神经协同过滤模型中,以增强模型对低曝光物品的探索能力;最后,在MovieLens-100K和MovieLens-1M数据集上进行实验,结果显示,与传统的神经协同过滤模型相比,该模型的曝光度提升了约60%,说明其能够有效地缓解曝光偏差,并提高推荐的准确性和公平性,进一步验证了该模型的有效性。In recommendation systems,strong exposure bias caused by sparse interaction data and uneven exposure tends to concentrate recommendations on highly exposed items,neglecting the potential value of low-exposure items,thus limiting user choices and diminishing user experience.To address this issue,this paper proposed a model that integrated neural collaborative filtering and the linear upper confidence bound(LinUCB)algorithm to mitigate exposure bias.Firstly,the model used neural collaborative filtering to analyze interaction data between users and items,learning their features and capturing latent preferences.Secondly,it introduced the LinUCB algorithm,embedding its generated reward feature into the neural collaborative filtering model to enhance the exploration capabilities for low-exposure items.Finally,experiments conducted on the Mo-vieLens-100K and MovieLens-1M datasets demonstrated that this model increased exposure by approximately 60%compared to traditional neural collaborative filtering models.This enhancement suggests that the proposed method effectively mitigates exposure bias and improves both the accuracy and fairness of recommendations,thereby validating the effectiveness of the model.

关 键 词：神经协同过滤 线性置信上界 曝光偏差 个性化推荐 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]