ENHANCING SEQUENTIAL RECOMMENDATION MODELING VIA ADVERSARIAL TRAINING

Recently, substantial progress has been made in the field of modeling sequential recommendation tasks through the application of deep neural networks. However, practical implementations of sequential deep learning models have been shown to be prone to representation degradation problems, which may leading to high semantic similarities among embeddings, and severely impact the recommendation performance and robustness. For alleviating this problem, in this paper, we propose a simple yet highly effective Adversarial Training mechanism for regularizing Sequential Recommendation models, namely ATSRec. In specific, we first conduct an empirical and theoretical study of this representation degradation problem. Then, we introduce adversarial perturbations to the item embedding layer, aiming to maximize the adversarial loss during model training. At last, theoretically, we show that our adversarial mechanism effectively encourages the diversity of the embedding vectors, helping to increase the robustness of models. Through extensive experiments on four public datasets and seven state-of-the-art models, we observed substantial improvements in both model overall performance and robustness with the help of ATSRec.