【作者】 陈俊彬；

【导师】 郝志峰；

【作者基本信息】 广东工业大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 用户显式行为的稀疏性及行为模式的多语义性是当前困扰用户偏好学习的现实问题。为了解决这一问题,一类重要方法是通过引入辅助信息(例如用户/项目属性、用户间关系等)来弥补显示行为的稀疏性,并通过异质信息网络刻画和建模行为模式的多语义性。然而,为了学习用户偏好,现有基于异质图的推荐方法主要通过预定义的模式(如元路径等)捕捉异质节点之间的高阶连接,这极度依赖于相关的领域知识;其次,对于用户-项目间关联模式的多语义性,还尚未有充分的研究。针对现有推荐方法面临的以上挑战,本论文提出了一种基于异质图的推荐方法,通过引入属性模体,其作为异质图中反复出现的高阶结构,用于揭露异质图中有意义的选择模式,从而辅助用户偏好的学习和推荐。本论文的主要工作是:(1)由于异质图中频繁出现的子结构往往能够体现用户的行为模式,尤其是选择模式(即用户如何选择项目),因此,区别于预定义,论文提出使用统计显著的子结构(即属性模体)来捕捉异质图中有意义的选择模式。发现异质图中这些用户的选择模式能够更好的捕获用户的偏好,进而辅助推荐决策。(2)为了有效的挖掘和表示异质图中节点间的多语义高阶关联信息,本论文通过构建基于模体的邻接矩阵来保存节点之间基于不同选择模式的高阶语义关联,并设计了一个基于图神经网络的框架来高效地建模多语义关联信息和协同过滤信号,以实现用户偏好学习和推荐。(3)为了更好的建模用户和项目关联模式的多语义性并将它们进行高效整合以准确的学习用户偏好。在上述框架的基础上,论文引入了邻居级别和语义级别注意力机制模块,这两个层次的注意力机制能够学习不同邻域以及关联模式的重要性并用于融合多语义信息,从而提高推荐效果。(4)本论文将提出的相关方法在三个真实数据集上进行了大量的实验,结果表明论文提出的模型皆优于近年来基于协同过滤以及基于异质图的相关基线方法。此外,通过对实验结果数据的分析,验证了使用属性模体捕捉用户选择模式的有效性。更多还原

【Abstract】 The sparsity of explicit user behaviors and the multi-semantics of behavioral patterns are the real problems that plague user preference learning.In order to solve this problem,an important method is to compensate for the sparsity of explicit behavior data by introducing auxiliary information(such as user / item attributes,relationship between users,etc.),and characterize and model the multi-semantics of behavior patterns through heterogeneous information networks(HIN).However,in order to learn user preferences,the existing HIN-based recommendation methods mainly capture the high-order connections between heterogeneous nodes through predefined patterns(such as meta-paths etc.),which is extremely dependent on the relevant domain knowledge;Secondly,the multi-semantics of user-item association patterns has not been fully studied.To address the above challenges faced by existing recommendation methods,this thesis proposes a novel HIN-based recommendation method.The thesis introduces attribute motifs,which are recurring higher-order substructures in HIN,to reveal meaningful selection patterns in HIN,so as to assist user preference learning and recommendation.The main work of this thesis is:(1)Because the frequent substructures in HIN can reflect the user’s behavior patterns,especially the selection patterns(i.e.how users choose items),different from predefined patterns,this thesis proposes to use statistically significant substructures(i.e.attribute motifs)to capture meaningful selection patterns in HIN.The discovery of user selection patterns in HIN can better capture users’ preferences and then assist in recommendation decisions.(2)In order to effectively mine and represent the multi semantic high-order association information between nodes in HIN,this thesis constructs a motif-based adjacency matrix to preserve the high-order semantic association between nodes based on different selection patterns,and designs a framework based on graph neural network to efficiently model the multi semantic association information and collaborative filtering signals,so as to realize user preference learning and recommendation.(3)In order to better model the multi-semantics for user and item association patterns and integrate them to accurately learn user preferences.Based on the above framework,the thesis introduces the neighbor-level and semantic-level attention mechanism modules.These two levels of attention mechanisms can learn the importance of different neighborhoods and association patterns,and can be used to fuse multi semantic information,so as to improve the recommendation performance.(4)Experiments on three datasets show that the method in this thesis is superior to the latest CF-based and HIN-based methods.In addition,analysis of experimental results also verifies the effectiveness of using attribute motifs to capture user’s selection patterns.更多还原