Knowledge graphs (KGs) have garnered significant attention in recommender systems as auxiliary information. Most existing studies consider an item as an entity of a KG and utilize graph neural networks to learn item representations. However, two challenges exist regarding these algorithms: 1) they provide recommended results but fail to explain the reason for which they are preferred by users; 2) user vector representations are concentrated in a small area, thus resulting in similar mass recommendations. In this study, we focus on learning fine-grained user preferences (LFUP) via user-item interactions and using KGs that can capture the reason for which users interact with items. Additionally, a personalized recommendation task is achieved by optimizing the distribution of users in the vector space. User preferences are modeled by using historical interaction items pertaining to users and important relations within the KG. Subsequently, information from two views is aggregated to reduce the semantic differences between them. Finally, user preferences are personalized by maximizing the spatial distance between various user representations via contrastive learning. Experiments on public datasets prove that LFUP significantly benefits user-preference modeling and personalized recommendations.