Since knowledge graphs (KGs) can effectively address the issues of sparsity and cold-start in collaborative filtering, they have been widely studied and used as auxiliary information in the field of recommendation systems. However, most existing KG-based recommendation methods primarily focus on leveraging the knowledge associations within the KG to represent users and items, while neglecting the latent signals inherent in the entities of the KG itself. Consequently, the learned embeddings fail to effectively represent the latent semantics of users and items in the vector space. In this paper, we propose a novel method named Relation-Fused Attention in Knowledge Graphs (RFA-KG) for Recommendation. This method introduces cross-compression networks to enhance the interactions among user-item interaction triples and incorporates a contrastive loss term to help the model learn from the information in the knowledge graph. Additionally, we design dynamic gate control mechanisms to dynamically perceive and adjust model weights, thereby adapting to different data characteristics and improving overall performance. Experimental results demonstrate that ours RFA-KG significantly outperforms several strong state-of-the-art baseline methods.