作为解决信息过载问题的有效方式，推荐系统已经被广泛地应用于新闻资讯、电子商务、社交媒体和个性化广告等领域。在推荐系统中，如何更好地利用用户行为数据挖掘用户和物品的有效特征并进行精准推荐一直是一个研究热点。近年来，由于超图神经网络在建模复杂数据关联时体现出了更强的表达能力和灵活性，一些学者开始使用超图神经网络来建模推荐系统中用户和物品之间的高阶关联信息，并取得了性能提升。然而，现有推荐算法仍存在一些不足。一方面，在建模用户和物品的关系时未能充分地考虑其交互时存在的不同意图；另一方面，仍未充分探索存在多种不同类型对象的推荐任务，如捆绑推荐。针对这两个问题，本文分别提出了基于解耦式双通道超图神经网络的协同过滤算法和基于多通道超图神经网络的捆绑推荐算法。在基于解耦式双通道超图神经网络的协同过滤算法中，我们首先将用户和物品的特征映射到不同的子空间中并在其中构建概率超图，其中不同的子空间表示不同的意图。随后，我们通过使用多个超图解耦模块使得不同意图下的超图结构和节点特征尽可能地具有差异性，从而使其分别关注于不同意图内的信息。该超图解耦模块包含意图感知的超图结构更新层、超图卷积层和一个关于更新后特征的距离相关性损失函数。通过在两个公开数据集上进行实验，我们验证了所提出的算法在协同过滤任务上的有效性。在基于多通道超图神经网络的捆绑推荐算法中，我们同时利用用户与物品的交互关系、用户与捆绑包的交互关系和物品与捆绑包的从属关系这三种不同类型的信息，分别在用户、物品及捆绑包这三个通道内构建超图。然后，我们通过在不同通道的超图结构上使用多个超图卷积模块来捕捉节点间的高阶关联信息，从而得到相应节点的特征表示。此外，我们进一步引入一个捆绑包引导的物品池化模块来聚合捆绑包内物品的特征，并以此补充捆绑包的特征信息从而提升捆绑推荐的性能。通过在两个公开数据集上进行实验，我们验证了所提出的算法在捆绑推荐任务上的有效性。

As an effective method to handle the problem of information overload, recommendation systems have been widely applied to various fields, such as news, e-commerce, social media, and personalized ads. In recommendation systems, it is always an important issue to make better use of user behavior data to mine the effective features of users and items and make accurate recommendations. In recent years, due to the capacity and flexibility of hypergraph neural networks in modeling complex data correlation, some works apply hypergraph neural networks to model high-order relationship information among users and items in recommendation systems, and achieve performance improvement. However, the existing recommendation algorithms still have some shortcomings. In the existing algorithms, the different intents of each user-item interaction are not fully considered when modeling user-item relationships. Besides, the existing algorithms have not fully explored recommendation tasks that have several different types of objects, such as, bundle recommendation. To tackle these two problems, we propose a collaborative filtering algorithm based on a disentangled dual-channel hypergraph neural network and a bundle recommendation algorithm based on a multi-channel hypergraph neural network separately.In the collaborative filtering algorithm based on a disentangled dual-channel hypergraph neural network, we firstly project the features of users and items to different subspaces and construct corresponding probabilistic hypergraphs, in which different subspaces represent different latent intents. Then, we employ multiple hypergraph disentangling modules to make the hypergraph structures and node embeddings of different intents independent of each other, to make them focus on the information of different intents. The hypergraph disentangling module is composed of an intent-aware hypergraph structure updating layer, a hypergraph convolutional layer, and a distance correlation loss function about the updated node features here. By conducting experiments on two public datasets, we verify the effectiveness of the proposed algorithm in the task of collaborative filtering.In the bundle recommendation algorithm based on a multi-channel hypergraph neural network, we utilize user-item interactions, user-bundle interactions, and item-bundle affiliations simultaneously to construct hypergraphs for the user, item, and bundle channels separately. Then, we apply multiple hypergraph convolution modules on the hypergraph structure of each channel to capture high-order correlation information among nodes, so as to get the corresponding features. Besides, we further introduce a bundle-guided item pooling module, which aggregates the features of items in the bundle to complement the feature of the bundle. By conducting experiments on two public datasets, we verify the effectiveness of the proposed algorithm in the task of bundle recommendation.