【作者】 刘杰；

【导师】 曾长清；

【作者基本信息】 南昌大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 现如今,我们处于数据量快速膨胀的信息时代,信息过载非常严重。我们平常在网上接受的信息量远超我们实际的需要,大量冗余的信息严重影响了我们获取有用信息。为了缓解信息过载对人们生活造成的影响,研究人员提出了很多值得称赞的解决方案,其中被广泛应用的方案是推荐系统。现在主流的推荐方法是协同过滤和基于内容做推荐,这些推荐方法难以捕捉用户的瞬时、短暂偏好。序列推荐系统解决了这类问题,通过对用户和项目之间交互序列的建模,能够很好地捕捉到用户随时间动态变化的偏好,从而进行更精准更符合用户近期偏好的推荐。推荐系统中的大部分数据存在图结构,适用于图神经网络,而近几年图神经网络在数据学习理解能力上的优异表现,吸引了众多研究人员在推荐系统模型中加入图神经网络。这些融合图神经网络的序列推荐研究虽然在实际推荐场景中取得了出色的成绩,但是仍然存在以下不足:(1)序列建图过程中相邻节点信息聚合时项目时间信息丢失问题。(2)只对用户目的单行为项目序列建模未发掘加入用户其余行为项目序列一起建模用于推荐的效果。(3)单一使用序列信息建模忽略了项目间相似属性可增强项目表达提升推荐效果。针对这些问题,本文在现有的使用图神经网络结构的序列推荐算法研究上做了如下三个创新:(1)提出了一种基于项目与推荐时刻时间间隔建图的序列建图方法,该方法增强了序列图中节点之间的边信息,保留了原序列中项目的时间信息。(2)提出了一个基于用户多个行为项目序列建模的图神经网络序列推荐模型,该模型发掘了不同行为对项目的影响,丰富了项目的嵌入表达。(3)提出了一个基于知识图谱增强的图神经网络序列推荐模型,该模型使用知识图谱作为辅助信息加入模型捕捉了不同项目属性间的联系,缓解了推荐系统的数据稀疏问题,提升了模型的推荐效果。更多还原

【Abstract】 Nowadays,we are in an information age where the amount of data is rapidly expanding,and information overload is very serious.The amount of information we receive on the internet far exceeds our actual needs,and a large amount of redundant information severely affects our ability to obtain useful information.In order to alleviate the impact of information overload on people’s lives,researchers have proposed many commendable solutions,among which the widely used solution is recommendation systems.The mainstream recommendation methods nowadays are collaborative filtering and content-based recommendation,which have difficulty capturing users’ instant and fleeting preferences.Sequential recommendation systems solve such problems by modeling the interaction sequences between users and items,which can capture users’ dynamically changing preferences over time,and thus provide more precise and relevant recommendations based on the users’ recent preferences.Most of the data in recommendation systems exist in graph structures,which make them suitable for graph neural networks.In recent years,the excellent performance of graph neural networks in learning and understanding data has attracted many researchers to incorporate them into recommendation system models.Although these sequential recommendation studies that integrate graph neural networks have achieved outstanding results in practical recommendation scenarios,there are still some shortcomings,including:(1)The loss of project sequence information when the information of adjacent nodes is aggregated in the process of sequence mapping.(2)Only model the item sequence for the user’s purpose single line without exploring the effect of adding the user’s other behavior item sequence to model together for recommendation.(3)The single use of sequence information modeling to make recommendations ignores the impact of similar attributes between items on the recommendation effect.In view of these problems,this paper has made the following three innovations in the existing research of sequence recommendation algorithm using graph neural network structure:(1)A sequence mapping method based on the time interval between items is proposed.This method enhances the edge information between nodes in the sequence diagram and retains the sequence information between items in the original sequence.(2)A graph neural network sequence model based on user’s multiple behavior item sequence modeling is proposed.This model explores the impact of different behaviors on the item and enriches the embedded expression of the item.(3)A graph neural network sequence model based on knowledge atlas enhancement is proposed.The model uses knowledge atlas as auxiliary information to join the model to capture the relationship between different project attributes,alleviate the data sparsity problem of the recommendation system,and improve the recommendation effect of the model.更多还原