【作者】 李浩；

【导师】 梁京章； 史水平；

【作者基本信息】 广西大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 在当今大数据时代的环境背景下,规模庞大的互联网数据信息无法得到有效的利用,“信息过载”日益加重。使用推荐系统可以有效地为用户建立模型,从而为用户过滤互联网信息,极大地降低“信息过载”带来的不利影响。传统的推荐系统在进行用户建模时,仅从用户的历史评分数据中提取信息,推测用户的偏好,从而为用户实现推荐。但随着互联网中数据量的不断增多,如网络购物平台中新注册用户数量和商品的种类数量日益增加,冷启动问题严重的同时也使得用户历史评分数据的稀疏性不断增大,传统推荐系统的推荐性能受到很大的影响。为解决以上问题,最常见的方法是引入其他辅助信息,而用户的文本评论信息代表着用户对物品的主观评价,含有大量的用户信息和物品信息,因此用户的历史评论信息被广泛地应用在各类推荐系统中,但如何从用户的历史评论文本中挖掘更多、更具个性化的用户和物品信息是当前各推荐系统亟待解决的问题。为获得更好的推荐性能,本文的主要研究内容如下:(1)针对冷启动问题,引入用户和物品的属性信息,利用用户、物品的属性相似度实现对新用户和新物品的推荐,缓解冷启动对推荐性能的影响。(2)为更好的从评论文本中提取用户和物品的特征,本文提出一种基于评论文本的个性化推荐算法,算法模型采用两个并行的卷积神经网络分别对用户和物品进行文本特征提取,同时在算法模型中引入三级注意力机制,分别从单词级别、语句级别、评论级别为用户和物品挖掘更具个性化的特征。目前大多数基于评论文本的推荐算法模型在进行文本特征提取过程中缺少用户与物品之间的交互特征,因此本文在使用两个并行卷积神经网络模型的同时,在模型中引入共同注意力网络,模拟用户与物品之间的交互从而挖掘更细粒度的用户和物品特征。(3)为验证本文算法模型在提升推荐效果上的有效性,在Amazon不同领域的五个数据集上设置了对比实验,实验结果显示本文算法在推荐效果上表现最优;最终为验证本文主要研究内容在提升推荐效果上的作用,本文设置了退化实验,实验结果显示本文研究内容皆有助于挖掘用户和物品的文本数据特征,从而达到提升推荐效果的作用。(4)以本文算法模型为系统核心设计并完成了个性化电影推荐系统,并通过设置系统测试验证了系统能够稳定运行。更多还原

【Abstract】 In the context of today’s big data era,the massive Internet data and information cannot be effectively utilized,and the adverse impact of "information overload" is increasingly aggravated.The recommendation system can effectively build models for users,thus filtering Internet information for users,and greatly alleviating the adverse effects of "information overload".Traditional recommendation systems only extract information from the user’s historical rating data when modelling the user,inferring the user’s preferences and making recommendations for the user.However,with the increasing amount of data in the Internet,such as the number of newly registered users and the variety of products in online shopping platforms,the cold start problem is serious and the sparsity of the user’s historical rating data is increasing,which greatly affects the recommendation performance of traditional recommendation systems.In order to solve the above problems,the most common method is to introduce other auxiliary information,and the user’s text review information represents the user’s subjective evaluation of the item,containing a large amount of user information and item information,so in recent years,the user’s historical review information is widely used in various recommendation systems,but how to mine more and more personalized user and item information from the user’s historical review text is the current various However,how to mine more and more personalized user and item information from the user’s historical comment text is an urgent problem for recommendation systems.In order to obtain better recommendation performance,the main research of this paper is as follows.1.Introducing the attribute information of users and items,using the attribute similarity of users and items to achieve recommendations for new users and new items,alleviating the impact of cold start on recommendation performance.2.The algorithm model uses two parallel convolutional neural networks to extract text features for users and items respectively,and introduces a three-level attention mechanism in the algorithm model to extract more personalized features for users and items from word level,statement level and comment level respectively.Most of the current recommendation algorithm models based on review text lack the interaction features between users and items in the text feature extraction process,so this paper uses two parallel convolutional neural network models and introduces a common attention network in the model to simulate the interaction between users and items in order to mine more finegrained user and item features.3.In order to verify the effectiveness of this paper’s algorithm model in improving the recommendation effect,this paper set up comparison experiments on five datasets from different areas of Amazon,and the experimental results show that this paper’s algorithm performs best in the recommendation effect;in order to verify the role of this paper’s main research content in improving the recommendation effect,this paper set up ablation experiments,and the experimental results show that this paper’s research content all help to explore the text data features of users and items,so as to achieve the role of improving the recommendation effect.4.A personalized movie recommendation system is designed and completed with the algorithm model as the core of the system,and the system can run stably through setting system test.更多还原