【作者】 曾志伟；

【导师】 陈健； 丁圣勇；

【作者基本信息】 华南理工大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 随着互联网技术的快速发展,网络负担日益加重,推荐系统能够帮助用户解决信息过载问题,为用户提供高效、精准地获取个性化信息的服务,因此受到学术界和工业界的广泛关注。目前大多数推荐模型只是基于用户与项目的交互关系进行建模,而忽略了用户属性和项目属性的辅助信息,以及忽略了用户与用户之间、项目与项目之间的隐藏信息。另外,由于图卷积网络结构的特点,当网络层数加深时,节点的特征会趋于相似,出现过平滑现象,导致模型的性能下降。针对属性表示和协作信号提取问题,我们提出了一种基于多信息融合的图卷积推荐方法。该模型包含了两个模块,分别是属性表示模块和协作提取模块。属性表示模块能够融合用户节点、项目节点的多种辅助属性信息,协作提取模块能学习用户-用户、项目-项目、用户-项目的隐藏信息,通过多信息图卷积将属性图和交互协作图的信息进行融合,从而对节点的特征属性及其交互关系更好地进行建模。与现有的先进方法相比,提出的方法在Movielens-1m数据集上Recall值和NDCG值分别提升了1.47%和1.12%,在Lastfm数据集上Recall值和NDCG值分别提升了1.45%和3.95%,充分验证了提出的方法能有效提升模型的推荐性能。针对邻居节点的权值分配和图卷积中的过平滑问题,我们提出一种基于残差连接的深层注意力图卷积推荐方法。设计了一个权重聚合模块,通过简要的注意力机制,学习不同邻居节点的贡献。设计了一个残差模块,在消息更新中引入残差结构,将节点的浅层特征输入深层网络,可以缓解图卷积中的过平滑问题,有效训练更深层数的网络模型。本文将网络模型深度设为5层,在Gowalla、Yelp2018、Amazon-book数据集上进行实验。与现有先进方法相比,Recall值分别提升了2.04%、3.39%和4.08%,NDCG值分别提升了1.41%、4.32%和4.88%,充分验证了提出的方法能有效缓解深层图卷积网络中的过平滑问题。更多还原

【Abstract】 With the rapid development of Internet technology,the network burden is increasing.Recommender System can help people solve the problem of information overload and provide users with efficient and accurate services to obtain personalized information.Therefore,it has attracted extensive attention from academia and industry.At present,most recommendation models are only based on the interaction between users and items,ignoring the auxiliary information of user attributes and item attributes,as well as the hidden information in user-user and item-item.In addition,due to the characteristics of graph convolution network structure,when the number of network layers is deepened,the characteristics of nodes will tend to be similar,resulting in over smoothing,resulting in the decline of the performance of the model.For the problem of attribute representation and cooperative signal extraction,we propose a graph convolution recommendation method based on multi information fusion.The model includes two modules: attribute representation module and collaborative extraction module.The attribute representation module can fuse a variety of auxiliary attribute information of user nodes and item nodes,and the collaboration extraction module can learn the hidden information of user-user,item-item and user-item,and fuse the information of attribute graph and interactive collaboration graph through multi information graph convolution network,so as to better model the characteristic attributes and interaction relationship of nodes.Compared with the existing advanced methods,the proposed method improves the Recall value and NDCG value by 1.47%and 1.12% on Movielens-1m data set and 1.45% and 3.95% on Lastfm data set respectively,which fully verifies that the proposed method can effectively improve the recommendation performance of the model.Aiming at the weight distribution of neighbor nodes and the over smoothing problem in graph convolution network,we propose a deep attention graph convolution recommendation method based on residual connection.A weight aggregation module is designed to learn the contribution of different neighbor nodes through a brief attention mechanism.A residual module is designed,the residual structure is introduced into the message update,and the shallow characteristics of nodes are input into the deep network,which can alleviate the over smoothing problem in graph convolution network and effectively train the deeper network model.On Gowalla,Yelp2018 and Amazon book data sets,the depth of the network model is set to 5 layers,the Recall value is increased by 2.04%,3.39% and 4.08% respectively,and the NDCG value is increased by 1.41%,4.32% and 4.88% respectively.It is fully verified that the proposed method can effectively alleviate the over smoothing problem in the deep graph convolution network.更多还原