【作者】 王森；

【导师】 李晓光；

【作者基本信息】 辽宁大学 ， 软件工程， 2023， 硕士

【摘要】 随着互联网技术的发展和数据爆炸式增长,渐渐的推荐在电商平台、音乐网站、社交网站等领域中变得越来越重要。早期的推荐方法主要是基于内容和协同过滤算法,然而在实际的推荐场景中,传统的基于内容和协同过滤的推荐算法存在冷启动、无法适应数据稀疏场景等问题。为了解决这些问题,近年来基于知识图谱的推荐方法逐渐成为研究热点,知识图谱是一种将实体、关系、属性等多种信息进行结构化表示的方式,能够将不同领域的知识进行融合,为推荐提供了更加全面和丰富的信息来源。但现有的基于知识图谱的推荐方法也会有不加区分地使用用户项目路径的问题,传递了不清楚的信息,并对可解释性产生了负面影响。针对以上这些问题,本文提出基于知识图谱路径可解释性的兴趣交叉融合推荐方法,简称ICFR(Interest Cross Fusion Recommendation),同样将知识图谱作为辅助信息的来源,利用知识图谱的结构信息和路径关系信息来增强推荐。ICFR采用深度端到端框架,并将嵌入任务和融合推荐任务结合,形成一种多任务特征学习的方法,起连接作用的兴趣交叉压缩单元实现潜在特征的自动共享,同时学习用户路径和知识图谱中的相似路径模式之间的高阶交互,证明了路径层面的交叉可以共享更多的潜在特征。本文的主要研究内容如下:首先,为了更好地表达用户的兴趣特征,本文使用基于知识图谱兴趣元路径的特征表达方法。具体而言,首先从生成的用户路径中提取出对应的元路径;然后,使用LSTM对路径进行建模,以得到路径上的语义特征表达;接下来,使用基于关系搜寻的双向BFS路径生成算法来搜索具有相同节点类型和关系模式的路径;最终,得到路径的语义特征表达用于增强用户兴趣特征偏好的可表达性。其次,提出了一种新的模型ICFR,即兴趣交叉融合推荐方法,以融合用户的兴趣偏好。该方法通过提取用户路径和知识图谱中具有相同关系模式的路径特征,并将两者进行交叉,实现了路径之间的高阶交互。相比于单纯的物品与实体的交叉,路径的交叉可以产生更多的特征交互,从而产生更多的辅助信息,也为推荐提供了可解释性。最后,将本文提出的算法模型与具有代表性的算法模型进行比较,在点击率(CTR)预测和top-k推荐这两个推荐场景下进行对比实验,并得到每个算法模型的AUC、ACC、Precision@K和Recall@K值,结果显示本文算法模型性能在电影、书籍、音乐推荐方面均优于其他对比算法模型。最后通过对比在电影数据集中降低训练集比例后在点击率预测场景中的AUC值,来测试在数据稀疏场景下各算法模型的表现,结果显示在用户-项目交互较为稀疏的情况下,本文所提出的算法模型仍能保持良好的性能。更多还原

【Abstract】 With the development of internet technology and explosive data growth,recommendation has gradually become increasingly important in fields such as e-commerce platforms,music websites,and social networking sites.The early recommendation methods are mainly content-based and collaborative filtering algorithms.However,in actual recommendation scenarios,the traditional content-based and collaborative filtering recommendation algorithms have some problems,such as cold start and failure to adapt to data sparse scenarios.In order to solve these problems,recommendation methods based on knowledge graphs have gradually become a research hotspot in recent years.Knowledge graph is a structured representation of entity,relationship,attribute and other information.It can integrate knowledge from different fields and provide more comprehensive and rich information sources for recommendation.However,the existing knowledge graph-based recommendation methods also have the problem of indiscriminate use of user project paths,conveying unclear information and having a negative impact on interpretability.In response to these issues,this article proposes an interest cross fusion recommendation method based on the interpretability of knowledge graph paths,abbreviated as ICFR(Interest Cross Fusion Recommendation).It also uses knowledge graph as a source of auxiliary information,utilizing the structural information and path relationship information of knowledge graph to enhance recommendation.ICFR adopts a deep end-to-end framework and combines embedding tasks with fusion recommendation tasks to form a multi task feature learning method.The interest cross compression unit plays a connecting role in achieving automatic sharing of potential features,while learning high-order interactions between user paths and similar path patterns in the knowledge graph,proving that path level intersections can share more potential features.The main research content of this article is as follows:Firstly,in order to better express the user’s interest features,this article uses a feature expression method based on the knowledge graph interest element path.Specifically,first extract the corresponding meta path from the generated user path.Then,LSTM is used to model the path to obtain semantic feature representations on the path.Next,use a bidirectional BFS path generation algorithm based on relationship search to search for paths with the same node type and relationship pattern.Finally,the semantic feature expression of the obtained path is used to enhance the expressibility of user interest feature preferences.Secondly,a new model called ICFR,namely the interest cross fusion recommendation method,was proposed to fuse users’ interest preferences.This method achieves high-order interaction between paths by extracting path features with the same relationship pattern in user paths and knowledge graphs,and intersecting the two.Compared to the simple intersection of items and entities,the intersection of paths can generate more feature interactions,thereby generating more auxiliary information and providing interpretability for recommendations.Finally,the algorithm model proposed in this article is compared with representative algorithm models,and comparative experiments are conducted in two recommendation scenarios: click through rate(CTR)prediction and top-k recommendation.The values of AUC,ACC,Precision@K and Recall@K of each algorithm model are obtained.The results show that the algorithm performance of the proposed model is superior to other comparison algorithm models in terms of movie,book and music recommendation.Finally,by comparing the AUC values in the click through rate prediction scenario after reducing the proportion of the training set in the movie dataset,the performance of each algorithm model in sparse data scenarios was tested.The results showed that the algorithm model proposed in this paper still maintained good performance even in situations where user project interaction was sparse.更多还原