Accurate preference modeling for sequential recommendation with adaptive denoising and attention

The fundamental objective of sequential recommendation is to meticulously analyze the user’s historical interaction behavior and accurately capture their preferences in order to predict the subsequent items that the user may be interested in. However, existing approaches predominantly focus on the interaction characteristics between users and items as well as the attributes of the items themselves, and tend to overlook the time-series dynamics of user preferences, including their longterm and short-term fluctuations. In addition, factors such as user misuse and system exposure bias may lead to noise in real-world interaction data, which poses significant challenges to the accuracy of the model. To address these challenges, we have designed an innovative model architecture that not only recognizes and corrects noise in the input initial item sequences, but also fuses users’ long-term and short-term preferences based on the corrected interaction sequences using a hierarchical attention mechanism to more realistically map users’ true preferences. In order to train our model to recognize and remove noise, we design two self-supervised learning mechanisms: deletion recovery and insertion removal. We conduct a comprehensive series of experiments on three real datasets and show that our model outperforms state-of-the-art sequential recommendation baseline models.