|
The negative item sampling strategies adopted in current recommendation models with implicit feedback have difficulty discriminating real negative examples from false negative ones (i.e., potentially positive ones). As a result, significant amounts of misleading information are used in the representation learning of users and items, resulting in slow convergence speeds and unsatisfactory recommendation results. To avoid the drawbacks of these approaches, we propose a novel model-agnostic framework, called the Affinity and Uniqueness Learning with Adaptive Data Augmentation (AUL-AD) framework, which does not require sampling negative user-item pairs from unobserved interactions. Specifically, we design an affinity and uniqueness learning objective in AUL-AD, which aims to encourage the similarity between positive-related users and items and the discrimination of each user (or item). Because there is a lack of supervised signals from inactive users and long-tailed items for affinity-uniqueness learning, we further designed a self-supervised task with an adaptive augmentation scheme based on user activity and item popularity. Extensive experiments on three real-world datasets demonstrate the superiority of our proposed framework, which can significantly improve the recommendation accuracy and efficiency. |
|
Keywords:;Recommender system; Data Augmentation; Self-supervised Learning |
|