三通道神经协同过滤算法  

Three-channel Neural Collaborative Filtering Algorithm

作　　者：周超 武友新[1] ZHOU Chao;WU You-xin(School of Information Engineering,Nanchang University,Nanchang 330000,China)

机构地区：[1]南昌大学信息工程学院,南昌330000

出　　处：《小型微型计算机系统》2022年第3期525-529,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61802085)资助;江西省科技计划项目(20151BBE50065)资助。

摘　　要：传统的神经协同过滤算法在隐式反馈数据集上对用户和项目建模,由于隐式反馈数据天然带有很强的噪音,这给模型的学习带来了挑战.为了缓解该问题,文中提出了一种基于三通道的神经协同过滤算法,该方法使用自编码器去挖掘用户和项目的特征向量,然后结合用户和项目的辅助信息一起通过多层感知机去学习特征向量不同维度之间的高阶交互关系,并将其与传统的神经协同过滤算法融合,以此来提高模型的泛化能力和命中率.此外,在隐式反馈数据集上进行负采样不易且采样结果会极大程度影响模型的表现,文中采用一种基于传统矩阵分解的概率负采样方法克服这个问题,提高了模型的鲁棒性.本文在公开数据集MovieLens上进行了大量实验,实验结果表明基于本文提出的算法比其他先进算法有更优的表现.The traditional neural collaborative filtering algorithms would model users and items on the implicit feedback data.The data only includes the interaction information between the users and the items.In other words,there is no user rating record of the item,which cannot reflect the users′real feelings for the item in a satisfied way.The model learned by this way would ignore the important information to make the recommendation contents are not always perfect.To alleviate the problem,a three-channel neural collaborative filtering algorithm is suggested.This method uses an auto-encoder to mine the feature vectors of users and items,and then combines the auxiliary information of users and items to learning high-level interaction relationships between the feature vectors in different dimensions through a multi-layer perceptron,and integrate it with the traditional neural collaborative filtering algorithm to improve the generalization ability and hit rate.Generally,sampling would significantly affect the model′s performance.Hence,a matrix factorization-based statistical negative sampling method is applied to ensure the model could learn enough information.Sufficient experiments have been tested on the public dataset,and the results have proofed that our method could achieve an impressive recommendation ability and strong robust ability than the other state-of-the-art methods.

关 键 词：神经协同过滤 负采样 隐式反馈 深度学习 推荐系统 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]