Robust Training of Sequential Recommender Systems with Missing Input Data

In the realm of sequential recommender systems, understanding users' preferences based on their past actions is paramount. Yet, the susceptibility of these models to input perturbations has limited their practicality. Addressing this, we present an innovative approach to mitigate the impact of missing input items, a challenge that has been overlooked. Our method involves a novel training process that anticipates data loss and employs an optimization loss to predict multiple future items. Extensive evaluations on diverse datasets and recommender models underscore its effectiveness. Notably, our approach enhances NDCG@10 by up to 18% with one missing item and an impressive 230% with five missing items, underscoring its substantial impact on system resilience and performance. This work sheds light on the intricate dynamics of sequential recommendation and offers a solution to real-world data limitations.