Aspect-based recommender systems (RSs) aim to provide accurate and explainable recommendations by leveraging fine-grained features from user reviews. However, existing methods do not utilize the accompanying sentiment information to enhance the discriminability of positive and negative preferences. To determine whether a user likes or dislikes an item, it is intuitive to assess the alignment or contradiction of item positive and negative aspects with the user’s preferred and rejected aspects. To realize this intuition, we propose a novel supervised contrastive learning approach that models relationships between ratings and aspect preferences for making recommendations. Our aspect representations are explicitly enriched with sentiments, capturing both semantic and sentimental aspects of user preferences. Additionally, we introduce a constraint to model the semantic relationship between observed and unobserved aspect preferences, enhancing recommendation accuracy. Extensive experiments demonstrate that our proposed framework consistently outperforms state-of-the-art RS methods not only in terms of accuracy but also in robustness to negative items.