Improving Sequential Recommendation with Global Item Transitions and Local Subsequences.

Many recent sequential recommendation methods have aimed to enhance the modeling of users' diverse interests to provide more accurate and varied recommendations. Despite great success, existing models possess a limitation in their capacity to solely perceive items within a user behavior sequence. Such an approach may potentially constrain their ability to effectively discover users' multiple interests. In this paper, we present a novel approach that addresses this gap by harnessing global item transitions to uncover users' potential interests, subsequently modeling them as local subsequences. To achieve this objective, we represent global item transitions as a directed graph, constructed by amalgamating behavior sequences from the training set. This graph‐based representation allows the design of a specialized module capable of aggregating global item transitional information and uncovering local subsequences specific to individual users. Consequently, downstream sequence models can focus on pertinent item transitions, disregarding unrelated ones. We evaluate the effectiveness of our method on four public data sets, outperforming existing alternatives or delivering competitive results. © 2023 Institute of Electrical Engineer of Japan and Wiley Periodicals LLC. [ABSTRACT FROM AUTHOR]