【作者】 王伟；

【导师】 贲晛烨； 李玉军；

【作者基本信息】 山东大学 ， 电子信息（专业学位）， 2025， 博士

【摘要】 作为缓解信息过载问题的有效工具,推荐系统利用数据分析和算法模型为用户提供个性化建议,其核心工作原理是通过挖掘用户行为数据和项目特征,预测用户可能感兴趣的内容或商品。通过分析用户的历史记录、偏好模式以及其他相关数据,推荐系统能够在海量信息中筛选出与用户需求最匹配的选项,从而帮助用户节省决策时间,提升使用体验。推荐系统在电子商务、流媒体服务和社交媒体等领域得到了广泛应用,极大地推动了用户参与度和商业收益的增长。推荐系统本质上是数据驱动的智能系统,其性能高度依赖于数据的规模、质量与安全。大规模的交互数据为训练复杂的推荐模型提供基础,而高质量、多样且安全的数据能显著提高推荐的准确度和用户满意度。然而,当前推荐系统领域围绕交互数据的利用与建模仍存在一些问题与不足,包括上下文数据缺失,数据增广质量低和用户数据隐私不安全。其中,上下文数据的缺失使推荐中的数据稀疏问题更加严重,阻碍了对用户特征的更完全建模;传统启发式数据增广策略难以生成准确反映物品转移关系的增广数据,影响了推荐系统预测用户下一交互的准确性;推荐系统中未经隐私化处理的交互数据容易造成用户隐私信息泄露,降低推荐系统的安全性和用户对推荐系统的信赖度。本论文分别针对上述三个问题,围绕推荐系统中的交互数据优化方法展开研究,提出三种对应的个性化推荐方法。文章的主要研究内容与贡献如下:(1)针对推荐中的上下文数据缺失问题,提出上下文感知的图神经协同过滤推荐方法,利用上下文数据优化用户特征建模。推荐系统中的用户-物品交互矩阵具有严重稀疏性,仅靠用户-物品交互数据难以准确建模用户的偏好特征。此外,现有研究中用户特征通常仅受物品和社交网络等长期因素的影响,忽略了时间等短期上下文因素,这也影响了对用户特征更完全的建模。为了解决这个问题,本论文针对推荐中的协同过滤任务,提出上下文感知的图神经网络模型,命名为CENTRIC。CENTRIC构建了一个包含用户-上下文协同过滤模块的多通道图神经网络,在用户特征中直接聚合邻居上下文特征,同时在上下文特征中聚合邻居用户特征。随着两种特征的交替更新,模型将有效捕捉到用户和上下文之间的多跳协作信号,丰富用户特征表示,进而缓解数据稀疏问题。此外,本文在网络输出层采用张量分解代替传统矩阵分解融合多通道特征,可以有效融合用户-物品-上下文交互,预测用户对物品的兴趣度得分。实验结果表明,本文引入缺失的上下文数据有效提升了图神经协同过滤推荐算法的性能。(2)针对推荐中的数据增广质量低问题,提出基于可学习数据增广的序列推荐方法,利用自监督学习的数据增广模块优化基于对比学习的序列推荐系统。对比学习方法常被用于缓解序列推荐中的数据稀疏问题,而数据增广是对比学习的核心,增广数据的质量直接影响对比学习和推荐系统的性能。然而,现有基于对比学习的序列推荐研究大多使用启发式或半启发式的增广方法,其过程往往具有大量随机性,难以生成符合用户真实交互意图的增广序列。不同于现有研究,本文训练了一个自监督学习的序列增广模块,该模块可以自动利用来自全局交互数据的自监督信息选择适当位置进行增广操作,生成的增广序列能够更准确反映物品转移关联和用户的真实意图。此外,本文进一步构建了由原始序列、可学习增广序列和随机增广序列组成的三元组,并设计了一个基于排序的三元对比损失,以增强来自可学习增广序列的正反馈,为推荐模型的参数优化提供更细致的对比信号。实验结果表明,所提出的可学习数据增广和三元对比学习都提升了序列推荐的性能。(3)针对推荐中的数据隐私不安全问题,提出基于协同混淆的隐私保护序列推荐方法,设计新颖的隐私保护机制优化序列推荐系统的数据安全。现有研究大多采用差分隐私或者联邦学习技术对推荐系统进行隐私保护,这通常会导致明显的性能下降或者对通信环境有较高要求。更重要的是,现有研究大多只适用于协同过滤任务,忽视了针对序列推荐的隐私保护。本论文提出了名为CLOUD的模型从不同于现有技术的视角实现针对序列推荐的隐私保护。对于一条目标序列,该模型根据其相似序列的共享表示对序列中的每一个物品执行保留、删除或者插入操作。此外,本文设计一个复制机制使来自相似序列的物品有更高的概率被插入到目标用户序列中。最后,修改后的交互序列取代原始序列,被输入到推荐系统中。由于修改后序列仅保留少部分真实物品,因此减少了用户真实交互数据泄露的风险。实验结果表明,CLOUD实现了最高66.57%的全局修改率,且推荐准确率持平先进的序列推荐模型。这证明CLOUD能够以很小的性能成本有效保护用户隐私,从而为推荐系统的隐私保护提供新的思路和技术方案。更多还原

【Abstract】 As an effective tool for mitigating information overload.recommender systems utilize data analytics and algorithmic models to provide users with personalized suggestions.Their core operating principle is to predict content or products that a user might be interested in by mining user behavior data and item characteristics.By analyzing users’historical records,preference patterns,and other relevant data,recommender systems are capable of filtering through vast amounts of information to identify options that best match the users’ needs,thereby saving decision-making time and enhancing user experience.Recommender systems have been widely applied in fields such as e-commerce,streaming services,and social media,significantly boosting user engagement and commercial revenue.Recommender systems are essentially data-driven intelligent systems whose performance heavily depends on the scale,quality,and security of the data.Large-scale interaction data provide the foundation for training complex recommendation models,while high-quality,diverse,and secure data can substantially improve recommendation accuracy and user satisfaction.However,there are still several issues and shortcomings in the current research on recommender systems regarding the utilization and modeling of interaction data,including missing contextual data,low-quality of data augmentation,and insecure user data privacy.Specifically,the lack of contextual data exacerbates the data sparsity problem in recommendations,which hinders the more complete modeling of user features;traditional heuristic data augmentation strategies struggle to generate augmented data that accurately reflect the transition relationships between items,thereby affecting the prediction accuracy of users’ next interactions;and interaction data in recommender systems that have not undergone privacy-preserving processing can easily lead to user privacy breaches,reducing both the security of the recommender systems and users’ trust in them.This thesis addresses the above three problems by investigating interaction data optimization methods for recommender systems and proposing three corresponding personalized recommendation methods.The main research contributions are summarized as follows:(1)To address the issue of missing contextual data in recommendations,a contextaware graph neural collaborative filtering method is proposed.which utilizes contextual data to optimize user feature modeling.The user-item interaction matrix in recommender systems is severely sparse,making it difficult to accurately model users’ preference features based solely on user-item interaction data.Moreover,existing studies typically consider that user features are influenced only by long-term factors such as items and social networks,while neglecting short-term contextual factors like time,which hampers a more complete modeling of user features.To address this,this thesis introduces a contextaware graph neural network model for collaborative filtering tasks in recommendations,named CENTRIC.CENTRIC constructs a multi-channel graph neural network that incorporates a user-context collaborative filtering module,directly aggregating neighboring context features into user features,while also aggregating neighboring user features into context features.With the alternating update of these two types of features,the model can effectively capture multi-hop collaborative signals between users and contexts,enrich the representation of user features,and thereby alleviate the data sparsity problem.In addition,tensor factorization is adopted in the network’s output layer to replace the traditional matrix factorization for fusing multi-channel features,which effectively fuses the user-item-context interactions and predicts the interest scores of users towards items.Experimental results indicate that the introduction of the missing user-context interaction data significantly enhances the performance of the graph neural collaborative filtering recommendation algorithm.(2)To address the issue of low-quality of data augmentation in recommendations,a sequential recommendation method based on learnable data augmentation is proposed,which employs a self-supervised data augmentation module to optimize the contrastive learning-based sequential recommendation system.Contrastive learning methods are often used to alleviate the data sparsity problem in sequential recommendation,and data augmentation is the core of contrastive learning,the quality of the augmented data directly affects the performance of both contrastive learning and the recommendation system.However,most existing contrastive learning-based sequential recommendation studies rely on heuristic or semi-heuristic augmentation methods,which are often highly random and struggle to generate augmented sequences that accurately reflect the users’ true interaction intents.Unlike previous studies,this thesis trains a self-supervised sequence augmentation module that automatically leverages self-supervised signals from global interaction data to select appropriate positions for augmentation.The generated augmented sequences can more accurately reflect the transition relationships between items and the users’ true intents.Furthermore,a triplet consisting of the raw sequence,the learnable augmented sequence,and a randomly augmented sequence is constructed,and a ranking-based triplet contrastive loss is designed to enhance the positive feedback from the learnable augmented sequence,thereby providing a more nuanced contrastive signal for optimizing the parameters of the recommendation model.Experimental results demonstrate that the proposed learnable data augmentation and triplet contrastive learning both improve the performance of sequential recommendation.(3)To address the issue of data privacy insecurity in recommendations,a privacypreserving sequential recommendation method based on collaborative confusion is proposed,which designs a novel privacy protection mechanism to enhance the data security of sequential recommendation systems.Most existing studies adopt differential privacy or federated learning techniques for privacy protection in recommender systems,which often lead to noticeable performance degradation or impose high requirements on communication environments.More importantly,existing studies are mostly applicable only to collaborative filtering tasks,overlooking privacy protection in sequential recommendation task.This thesis proposes a model named CLOUD to achieve privacy protection for sequential recommendation from a perspective different from existing technologies.For a target sequence,the model performs ’keep’,’delete’,or ’insert’ operations on each item in the sequence based on a shared representation of its similar sequences.In addition,a copying mechanism is designed to make items from similar sequences have a higher probability of being inserted into the target user sequence.Finally,the modified interaction sequence replaces the raw sequence and is input into the recommendation system.Since the modified sequence retains only a small portion of the real items,the risk of exposing users’ actual interaction data is reduced.Experimental results indicate that CLOUD achieves a global modification rate of up to 66.57%while maintaining recommendation accuracy comparable to state-of-the-art sequential recommendation models.This demonstrates that CLOUD can effectively protect user privacy at a minimal performance cost,thus offering new ideas and technical solutions for privacy protection in recommender systems.更多还原