RACR: Robust and Accurate Unlabeled Classification for Recommendation

Recommendation systems typically rely on user interaction data to deliver personalized recommendations. However, since users are exposed to only a limited set of items, the absence of observed interactions does not necessarily imply negative preferences, leading to exposure bias. Existing research methods often struggle to accurately distinguish potential positive and negative samples, which undermines the system’s recommendation performance. To address this challenge, we propose a recommendation model based on a dual training strategy, named RACR, which comprises the Robust Classifier Construction (RCC) module and the Contrast-Enhance Retraining (CER) module. Firstly, the RCC module employs a positive-unlabeled (PU) learning mechanism to construct a robust classifier and integrates a perturbation-resistant optimization strategy to accurately identify reliable negative samples within unexposed data. Secondly, the CER module utilizes multi-kernel similarity computations and contrastive loss on positive and reliably determined negative samples, enhancing the model’s capability to capture high-level latent associations between users and items in the embedding space, thereby improving the accuracy of recommendations. Our empirical study on several datasets demonstrates that RACR significantly outperforms the state-of-the-art models.