Knowledge graph-based recommendation systems face significant challenges from long-tail distributions and insufficient semantic information extraction. We propose KGRGA (Knowledge Graph Recommendation with Graph enhancement and Adaptive propagation), a novel framework addressing these limitations through three key innovations. First, we introduce GRADE (Graph contrastive learning for Degree Bias), which generates enhanced interaction graphs that rebalance attention between popular and tail entities. Second, we develop an Adaptive Light Graph Convolution Network (ALGCN) that dynamically adjusts each node’s information reception field by computing node-specific propagation probabilities, effectively mitigating popularity bias. Third, we employ an enhanced PathNet to capture rich semantic and structural knowledge from the knowledge graph. Extensive experiments on Yelp2018, Amazon-Book, and MIND datasets demonstrate that KGRGA significantly outperforms state-of-the-art baselines, with particular effectiveness in handling long-tail distributions and complex semantic relationships.