Multi-behavior session-based recommendation aims to predict the next item, such as a location-based service (LBS) or a product, to be interacted by a specific behavior type (e.g., buy or click) in a session involving multiple types of behaviors. State-of-the-art methods generally model multi-behavior dependencies in item-level, but ignore the potential of discovering useful patterns of multi-behavior transition through feature-level representation learning. Besides, sequential and non-sequential patterns should be properly fused in session modeling to capture dynamic interests within the session. To this end, this paper proposes a Graph Neural Network based Hybrid Model GNNH, which enables feature-level deeper representations of multi-behavior interaction sequences for session-based recommendation. Specifically, we first construct multi-relational item graph (MRIG) and feature graph (MRFG) based on session sequences. On top of the MRIG and MRFG, our model takes advantage of GNN to capture item and feature representations, such that global item-to-item and feature-to-feature relations are fully preserved. Afterwards, each multi-behavior session is modeled by a seamless fusion of interacted item and feature representations, where self-attention and mean-pooling are used to obtain sequential and non-sequential patterns simultaneously. Experiments on two real datasets show that the GNNH model significantly outperforms the state-of-the-art methods.

基于多行为会话的推荐旨在预测下一个项目，例如基于位置的服务 (LBS) 或产品，在涉及多种类型的会话中通过特定行为类型（例如，购买或点击）进行交互。行为。最先进的方法通常在项目级别对多行为依赖性进行建模，但忽略了通过特征级别表示学习发现多行为转换的有用模式的潜力。此外，在会话建模中应该适当地融合顺序和非顺序模式，以捕捉会话中的动态兴趣。为此，本文提出了一种基于图神经网络的混合模型 GNNH，它能够对基于会话的推荐的多行为交互序列进行特征级更深层次的表示。具体来说，我们首先基于会话序列构建多关系项目图（MRIG）和特征图（MRFG）。在 MRIG 和 MRFG 之上，我们的模型利用 GNN 来捕获项目和特征表示，从而完全保留全局项目到项目和特征到特征的关系。之后，每个多行为会话都通过交互项目和特征表示的无缝融合进行建模，其中使用自注意力和均值池同时获得序列和非序列模式。在两个真实数据集上的实验表明，GNNH 模型明显优于最先进的方法。这样全局的 item-to-item 和 feature-to-feature 关系被完全保留。之后，每个多行为会话都通过交互项目和特征表示的无缝融合进行建模，其中使用自注意力和均值池同时获得序列和非序列模式。在两个真实数据集上的实验表明，GNNH 模型明显优于最先进的方法。这样全局的 item-to-item 和 feature-to-feature 关系被完全保留。之后，每个多行为会话都通过交互项目和特征表示的无缝融合进行建模，其中使用自注意力和均值池同时获得序列和非序列模式。在两个真实数据集上的实验表明，GNNH 模型明显优于最先进的方法。