基于样本选择的自学习去噪鲁棒模型  

Robust denoising model based on self-learning with sample selection

作　　者：王魁 姚迅[1] 杨捷 WANG Kui;YAO Xun;YANG Jie(School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan Hubei 430200,China;School of Computer Science,University of Wollongong,Wollongong 2500,Australia)

机构地区：[1]武汉纺织大学计算机与人工智能学院,武汉430020 [2]伍伦贡大学计算机学院,澳大利亚伍伦贡2500

出　　处：《计算机应用》2024年第S01期77-82,共6页journal of Computer Applications

摘　　要：隐式反馈数据能有效缓解推荐系统训练数据稀疏的问题,但它很难反馈真实的用户-项目偏好,导致数据中会包含较多噪声数据,从而影响推荐模型的性能。近年提出的一些去噪方法在一定程度上增强了推荐模型的鲁棒性,但这些方法基本都从数据的特性着手,通过获取额外辅助数据达到去噪的目的,然而辅助数据的获取困难,且难以泛化。因此,从推荐模型出发,提出一种基于样本选择的自学习依赖值评估器鲁棒模型对隐式反馈数据去噪。该鲁棒模型利用推荐模型性能对隐式反馈数据的依赖性,通过一种辅助梯度策略算法实现依赖值评估器的自主更新,可有效筛选依赖样本,并利用依赖样本训练推荐模型达到提升推荐性能的目的。在3个具有代表性的隐式反馈数据集和4种推荐模型上进行大量实验。实验结果表明,使用常见的归一化折损累计增益(NDCG@K)评估得分衡量推荐性能,在泛用的隐式反馈数据集Adressa、Amazon-book和Yelp上,与含噪声训练相比,所提鲁棒模型能分别提升NeuMF(Neural Matrix Factorization)推荐模型性能0.15、0.78和0.35个百分点;分别提升GMF(Generalized Matrix Factorization)推荐模型性能1.53、0.93和0.46个百分点。在相同实验评估下,使用数据集Adressa和Yelp,该鲁棒模型分别提升CDAE(Collaborative Denoising Auto-Encoders)推荐模型性能1.58和0.34个百分点;分别提升LightGCN(Light Graph Convolutional Network)推荐模型性能1.82和0.15个百分点。Implicit feedback data can effectively alleviate the problem of sparse training data for recommender systems,as implicit feedback data is difficult to feed back the real user-item preferences,resulting in the inclusion of more noisy data,which will affect the performance of the recommendation model.In recent years,many denoising methods have been proposed in order to enhance the robustness of recommendation models,however,these methods primarily focus on data features and rely on obtaining additional auxiliary data for denoising,which is challenging to generalize.Considering the characteristics of the recommendation model itself,a self-learning reliance value evaluator robust model based on sample selection was designed for denoising the implicit feedback data of the recommendation model.In the proposed robust model,based on the dependence of recommendation model performance on feedback data,an auxiliary gradient strategy algorithm was adopted to realized the autonomous update.The dependent samples clould be effectively filtered,and ultimately used to train the recommendation model to improve the recommendation performance.Extensive experiments were designed on three representative implicit feedback datasets and four recommendation models.Experimental results demonstrate that,using the common Normalized Discounted Cummulative Gain(NDCG@K)evaluation score to measure recommendation performance on the generalized implicit feedback datasets Adressa,Amazon-book and Yelp,the robust model improves the performance of the NeuMF(Neural Matrix Factorization)recommendation model by 0.15,0.78,and 0.35 percentage points,respectively;and improves the performance of GMF(Generalized Matrix Factorization)recommendation model by 1.53,0.93 and 0.46 percentage points,respectively,when compared to noise-laden training.On the datasets Adressa and Yelp,the robust model improves the performance of CDAE(Collaborative Denoising Auto-Encoders)recommendation model by 1.58 and 0.34 percentage points respectively;and improves the performance o

关 键 词：推荐系统 自学习 鲁棒性 隐式反馈 去噪 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]