Attribute and Context-Aware Multi-Behavior Model for Unique-Item Recommendation

In the context of sequential recommendation, incorporating auxiliary information has consistently shown improvements in several scenarios. Some models focus on integrating item and user features, other approaches include context information. A notable area of growth is the multi-behavior recommendation, which considers the different user’s behaviors to indicate their preferences. Current methods rely on graph-based models while other sequential multi-behavior models use transformers. However, none of these models take items or user features into consideration which causes a limited understanding of user preferences through different actions and item characteristics. In this paper, we propose an Attribute and Context-aware Multi-Behavior model (ACMB) for unique item recommendation. This not only accounts for users’ varying behaviors but also integrates the relevant attributes of items to enhance the understanding of user preferences. ACMB encodes the items with the respective attributes then it applies a hierarchical attention over the different behaviors separately, followed by attention across the entire input sequence to generate a comprehensive deep sequence representation. Extensive experiments on real-world Volkswagen Financial Services (VWFS) dataset demonstrate the significance of our proposed model over the current state-of-the-art attribute-aware sequential recommendation methods.