在信息技术飞速发展的大数据时代，信息过载问题成为了人们快速获取目标资源的巨大障碍。个性化推荐系统通过信息过滤给用户推荐合适的信息资源，在一定程度上缓解了信息过载问题。捆绑推荐作为传统单物品推荐的升级方案，可以一次性给用户推荐某个主题的多个物品，在电子商务、在线音乐和在线阅读等平台得到了广泛应用。服务提供者可以通过捆绑推荐提高推荐效率以获得更大的经济效益，用户也可以通过捆绑的形式降低逐个进行物品选择的时间成本。 为了提高捆绑推荐的性能表现，研究者在将图神经网络应用于捆绑推荐上进行了多方面的尝试。目前，“双视角学习+对比学习”的基本框架在捆绑推荐任务上取得了显著性能提升，但是聚合函数和图构建这两个将图神经网络应用于捆绑推荐的关键问题还未得到充分探索， 因此本文针对这两个问题进行了研究。 在图神经网络的聚合函数上，本文提出了在捆绑-物品从属关系上基于流行度的聚合函数，通过引入物品流行度因素来区分捆绑内包含物品的贡献差异性。同时在特定捆绑角度进行贝叶斯个性化排序，使模型对训练数据集的利用更加充分。在真实数据集的实验表明，该模型比最优基线模型取得了一定程度的性能提升。 在图神经网络的图构建上，本文提出通过用户-物品交互数据和捆绑-物品从属数据来推导一个假想的用户-捆绑交互图，通过这个推导图在图构建方面隐去了物品节点并将其应用于下游的捆绑推荐任务。在真实数据集的实验验证了方案的有效性，在部分特定领域的数据集上的性能相比最优基线模型取得了大幅度提升。

In the big data era with the rapid development of information technology, information overload has become a major obstacle for people to quickly obtain target resources. Personalized recommendation system recommends suitable information resources to users through information filtering, which has partially alleviated the problem of information overload. Bundled recommendation, as an upgraded solution for traditional single-item recommendation, can recommend multiple items on a particular topic to users at once and has been widely used in platforms such as e-commerce, online music, and online reading. Service providers can improve recommendation efficiency and achieve greater economic benefits through bundle recommendation, while users can get bundles to reduce the time cost of selecting items one by one. To improve the performance of bundle recommendation, researchers have made various attempts to apply graph neural networks to bundle recommendation. Currently, the basic framework of "dual-view learning + contrastive learning" has achieved significant performance improvement in bundle recommendation tasks, but the key issues of applying graph neural networks to bundle recommendation such as aggregation functions and graph construction have not been fully explored. Therefore, this paper conducts research on these two issues. In terms of aggregation functions in graph neural networks, this paper proposes a popularity-based aggregation function for bundle-item affiliations, distinguishing contributions of items in a bundle by introducing the item popularity. Meanwhile, bayesian personalized ranking is performed from bundle-specific view for better utilization of the training dataset. Experiments on real datasets show that the proposed model achieves a certain degree of performance improvement over the SOTA baseline. In terms of graph construction in graph neural networks, this paper proposes to derive a hypothetical user-bundle interaction graph with user-item interaction data and bundle-item affiliation data, and to hide item nodes in graph construction through the derived graph which is applied to downstream bundle recommendation tasks. The effectiveness of the proposed method is verified through experiments on real datasets and the proposed model achieves significant improvements compared to SOTA baseline on datasets of certain domain.