【作者】 赵清泉；

【导师】 王琪；

【作者基本信息】 吉林大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 在现代社会中,随着信息日益增长,信息过载成为人们常面临的问题。为了解决这一问题,推荐系统应运而生。推荐系统旨在通过分析用户的历史交互数据来预测其兴趣,并据此提供个性化推荐。这种方法在多种应用场景中被广泛采用,极大地减轻了信息过载的问题。然而,传统的推荐系统主要侧重于分析用户在单一领域内的行为数据,这种方法在实际应用中面临着数据稀疏性和冷启动问题,尤其是在新用户或新商品加入时,由于缺乏足够的交互数据,推荐的准确度和效率大大降低。跨域推荐系统通过引入辅助域的信息,利用不同领域间的相关性来增强推荐系统的性能,从而有效缓解单一领域推荐系统的数据稀疏性和冷启动问题。本文深入探讨了跨域推荐系统的两种主要方法:基于混合共享实体和基于映射关系的方法。基于混合共享实体类的跨域推荐系统有效地利用跨域的共享实体数据,充分利用深度学习技术来提取和学习跨域共享实体的特征表示,以更好地理解用户在不同领域的行为模式,克服不同领域数据稀疏性,捕获复杂的非线性关系和高维特征交互以充分的整合来自不同领域的信息。本文从这个角度出发,提出了基于Transformer的跨域推荐算法(Cross-Domain Recommendation System Based on Graphical and Transformer,GTCDR)。首先,借助图嵌入技术将来自不同领域的实体及其交互表示为嵌入向量,使得不同领域之间的信息能够在同一个向量空间中进行整合,便于模型理解和利用这些信息并进行推荐。然后得益于Transformer强大的特征提取能力和信息融合能力,算法可有效整合不同领域以及本领域的相关信息,从而提升领域推荐的准确性。基于映射的跨域推荐系统其实现方式一般通过找到两个领域之间的共通点或链接来实现。跨域推荐系统中领域是由用户与物品的交互所构成的,领域之中用户之间的个性不同,因而每个用户的相应特征也各不相同,用户个性特征与所在领域信息相结合所构成的本领域内用户的特征更增加了此类跨域信息迁移的复杂性,导致此类跨域推荐系统在进行跨域信息传输时很难精准的一步传输到位。针对以上所述问题,本文提出了基于对比学习抽取用户可迁移特征的跨域推荐算法(Contrastive Learning for Extracting Transferable User Profiles in Cross-Domain Recommendation System,CLUPCDR)。本算法通过对比学习来最大化两个领域中重叠用户的关联度和最小化非重叠用户的关联度,帮助模型捕捉跨域数据中更丰富、更具区分性的特征,以学习到可迁移的特征表示。两个算法分别是基于混合共享实体类的跨域推荐系统和基于映射的跨域推荐系统类别的算法改进,是分别针对于不同的应用场景的跨域推荐算法。然后,本文针对这两个算法进行了广泛的实验,覆盖了多个数据集。实验结果表明GTCDR能够发现不同领域之间的潜在联系,有效整合本领域及不同领域信息,CLUPCDR则通过对比学习,使模型学习到更加丰富和区分度高的可迁移特征表示,得到了更好的推荐效果。更多还原

【Abstract】 In modern society,with the ever-increasing volume of information,information overload has become a common problem that people face.To address this issue,recommendation systems have emerged,aiming to predict users’ interests by analyzing their historical interaction data and thus provide personalized recommendations.This approach has been widely adopted in various application scenarios,effectively mitigating the problem of information overload.However,traditional recommendation systems primarily focus on analyzing users’ behavior data within a single domain,which faces challenges such as data sparsity and the cold start problem,especially when new users or new items are introduced.Due to the lack of sufficient interaction data,the accuracy and efficiency of recommendations significantly decrease.Crossdomain recommendation systems,by introducing auxiliary domain information,can utilize the correlations between different domains to enhance the performance of recommendation systems,effectively solving the issues of data sparsity and cold start in single-domain recommendation systems.This paper delves into two main approaches of cross-domain recommendation systems: those based on hybrid shared entities and mapping relationships.Research on cross-domain recommendation systems based on hybrid shared entities mainly focuses on overcoming data sparsity across different domains,effectively utilizing cross-domain shared entity data,and fully leveraging deep learning techniques to extract and learn advanced feature representations of cross-domain shared entities.This aims to better understand users’ behavior patterns across different domains,capture complex nonlinear relationships and high-dimensional feature interactions,and fully integrate information from different domains.From this perspective,this paper proposes a cross-domain recommendation algorithm based on Graph and Transformer(GTCDR).Initially,by using heterogeneous graphs and graph embedding technology,entities and their interactions from different domains are represented as embedding vectors,allowing the information from different domains to be integrated into the same vector space,facilitating the model’s understanding and utilization of this information for recommendations.Then,thanks to the Transformer’s powerful feature extraction and information fusion capabilities,the algorithm can effectively integrate relevant information from different domains and within the same domain,thereby enhancing the accuracy and diversity of recommendations across the two domains.The implementation of mapping-based cross-domain recommendation systems generally involves finding common points or links between two domains.In cross-domain recommendation systems,domains are composed of interactions between users and items.With users in a domain having different personal characteristics,each user’s corresponding features also vary.Combining users’ personal characteristics with domain information to form indomain user features adds complexity to the transfer of cross-domain information,making it challenging for such cross-domain recommendation systems to accurately transfer information in one step.Addressing the aforementioned issues,this paper proposes a cross-domain recommendation algorithm based on contrastive learning for extracting transferable user profiles(CLUPCDR).This algorithm utilizes contrastive learning to maximize the affinity of the shared parts between two domains and minimize the affinity of the non-shared parts,helping the model capture richer,more distinctive features in cross-domain data to learn transferable feature representations.The two algorithms are improvements upon cross-domain recommendation systems based on hybrid shared entities and mapping-based cross-domain recommendation systems,respectively,and are designed for different application scenarios in cross-domain recommendation systems.Extensive experiments conducted on multiple datasets show that GTCDR can discover hidden connections between different domains,effectively integrate information within and across different domains,while CLUPCDR,through contrastive learning,enables the model to learn richer and more distinct transferable feature representations,resulting in better recommendation performance.更多还原