Recommendation Systems with Non-stationary Transformer

Recommendation systems rely on an accurate user model to understand users’ needs to make a personal recommendation. Traditional user modeling uses users’ past behaviors during a “supply-meets-demand” interaction. This approach failed to capture the dynamic and emergence of new items and the shifting of user interests. The recommendation systems, built based on this user model trap users in their previous interests and make recommendations without counting their interest shift. We propose a new approach that integrates a non-stationary transformer into a recommendation system to capture the temporal dynamics of supplies and shifting user interests. Our experiments demonstrate the framework’s superiority over benchmark models. The empirical results confirm the efficacy of our proposed framework and significant performance enhancements for recommendations.