Exploring the Side-Information Fusion for Sequential Recommendation

Side information fusion for sequential recommendation aims to mitigate the data sparsity problems by leveraging the additional knowledge besides item ID. While most state-of-the-art methods devised elaborate fusion methods to incorporate side-information, they overlooked that there are distinct characteristics of the side-information, which can be grouped into two types: item attribute (e.g., category and brand) and user behavior (e.g., position and rating). In this paper, we argue that attribute information and behavior information are fundamentally different in relation to the item. The former is inherent to the item, whereas the latter is not. Based on this intuition, we systematically analyzed the previous fusion approach and introduced a comprehensive framework for two types of side information. Finally, we devise self-supervised objectives fitting for each type of side-information in a multi-task training scheme. To validate the effectiveness of our proposed method, we conduct experiments across various domains.