A survey on multi-behavior sequential recommendation

People usually have the explicit or implicit desire to get the information they need and are most interested in from massive information, which has led to the creation of personalized recommender systems. Recommender systems are set up to address the issue of information overload in traditional information retrieval systems such as search engines, and have been a significant area of research focusing on recommending information that is of most interest to users. There is a sequential nature to the behaviors of a person interacting with a system, such as examining one item of clothing before examining others. The problem of taking this sequential nature into account to uncover users' interest dynamics in delivering recommendations is known as sequential recommendation (SR). The traditional SR problem merely focuses on a single behavior type of the users, while in real-world scenarios, users tend to engage in multiple types of behaviors, such as examining and adding clothes to the cart before purchasing them. The introduction of multiple behavior types can uncover users' behavior patterns more comprehensively, leading to the proposal of multi-behavior sequential recommendation (MBSR). MBSR considers both sequentiality and heterogeneity of user behaviors, which can achieve state-of-the-art recommendation performance through suitable modeling. Currently, there are some related studies for MBSR, and to the best of our knowledge, there is no related review to introduce and categorize these MBSR studies. Hence, this survey aims to shed light on MBSR, which is a relatively new and worthy direction for in-depth research. First, we introduce MBSR in detail, including its problem definition, application scenarios, and challenges faced. Second, we detail the classification of MBSR methods, including traditional methods and deep learning-based methods, where the former contain neighborhood-based methods and matrix factorization-based methods, and the latter can be classified into different learning architectures based on recurrent neural network (RNN), graph neural network (GNN), Transformer, and generic architectures, as well as architectures that integrate hybrid techniques. In each method, we present related studies from the data perspective and the modeling perspective, analyzing the strengths, weaknesses, and features of these studies, and further conduct experiments on two real-world datasets with classical and recent studies on different methods to show the difference in recommendation performance of these methods. Finally, we discuss some promising future research directions to address the challenges and improve the current status of MBSR.