【作者】 张慧；

【导师】 陈可佳；

【作者基本信息】 南京邮电大学 ， 计算机应用技术， 2021， 硕士

【摘要】 传统的推荐方法主要是抽取用户或项目的有效特征并利用矩阵分解技术学习用户和项目的隐向量矩阵。这类方法需要用户的显式反馈以及辅助特征,难以扩展至大规模数据,同时会因数据稀疏出现冷启动问题。由于推荐系统中用户和项目的交互数据本质上也可以抽象为非欧几里得空间的图(或网络),因此基于图学习的推荐系统得到了极大的关注。网络表示学习能够自动学习一个将节点映射到保留结构信息的低维稠密空间的函数,将其应用到推荐任务上可以获得用户和项目的表示。此类方法目前仍然存在以下挑战:(1)大多数研究基于用户的显式反馈构建网络,但是显式反馈较难获得从而导致交互数据过于稀疏;(2)大多方法忽视了用户偏好随时间演变的特性,将交互数据构建为一个静态网络。针对上述问题,本文系统性地进行了以下工作。(1)本文利用用户-项目之间大量的多类型隐式反馈构建多行为网络,提出了Ri Do TA模型。首先,将多行为网络分为一个基网络、多个源网络和一个目标网络;然后,基于随机游走联合优化来学习节点在各个网络中的嵌入表示;接着,为了学习不同的隐式反馈行为对目标行为(显式反馈)的影响,利用注意力机制自适应地学习各个源网络的权重;最后,将学习到的用户和项目的向量拼接,利用多层感知机学习用户和项目之间复杂的非线性关系并进行预测。(2)本文为了学习随时间演变的用户偏好,将动态网络表示学习融入推荐系统,提出了DBNRec模型。首先,利用用户-项目交互数据构建动态二分网络并且定义时序边权重;其次,利用图卷积网络聚合网络中两种类型节点的一阶邻居和高阶邻居信息,通过信息构建和信息传播更新节点的嵌入表示,同时考虑二分网络中节点的异质性,在两类节点嵌入矩阵相乘时引入了转换矩阵;最后,将用户和项目的节点表示进行拼接送入多层感知机获得预测得分。更多还原

【Abstract】 The traditional recommendation methods mainly use matrix factorization technology to extract the effective features of users.This type of methods requires lots of explicit feedback from users,and is difficult to scale to complex and large-scale data.Besides,cold start problems may occur due to sparse data.Since the interaction data between users and items can essentially be abstracted as a graph(or network)in a non-Euclidean space,recommendation methods based on graph learning have received great attention.Network representation learning aims to obtain the embedding of nodes which retains structural information and semantic information,and can be applied to recommendation tasks.This type of methods currently still has the following challenges:(1)Lots of studies build networks based on explicit feedback of users.However,this kind of data is hard to obtain,resulting in data sparsity problem;(2)Many methods ignore the evolution of user preferences over time.The interactive data between users and items is built as a static network.To handle the above challenges,this article systematically performed the following work.(1)This paper uses a large number of multiple types of implicit feedback between users and items to construct a multi-behavior network,and proposes the Ri Do TA model.First,the multi-behavior network is separated into a base network,multiple source networks and a target network;then,this model jointly optimizes the embedding representation of nodes in each network based on random walk strategy;then,so as to learn the impact of different implicit feedback behaviors on target behavior(explicit feedback),attention mechanism is applied to adaptively obtain the importance of each source network;finally,the learned embedding of users and items is concatenated as the input of multi-layer perceptron to learn the complex nonlinear relationship and predict the preference score between users and items.(2)To model user preferences that evolve over time,this paper integrates dynamic network methods into recommendation tasks and proposes the DBNRec model.First,DBNRec uses user-item interaction data to construct a dynamic bipartite network and defines temporal edge weights;second,graph convolutional networks are used to aggregate the first-order and high-order neighbor information of two types of nodes in the network.Embedding of nodes is updated through information construction and information dissemination.Considering the heterogeneity of the nodes in the bipartite network,a conversion matrix is introduced when the two types of node embedding matrices are multiplied;finally,the embeddings of users and items are spliced into the multi-layer perceptron for predictions.更多还原