Personalized Federated Learning (PFL) offers a privacy-preserving way to model user preferences. However, current methods often yield suboptimal models that struggle to predict user behavior in complex scenarios, reducing service accuracy and reliability. To address insufficient personalization and embedding bias in recommendation prediction, we propose Federated Recommendation with Composite Aggregation and Additive Personalization (FedCAAP) which introduces a dual-view modeling mechanism. It captures global user common features through a global view and individual perceptual differences through a local personalized view. Dual regularization ensures Information complementarity, enabling personalized recommendations. Additionally, a composite aggregation method is integrated into FL, aggregating both similar and complementary clients to update trained and untrained embeddings. Extensive experiments validate the effectiveness of the proposed method. FedCAAP outperforms the baseline by 2.9% to 3.37% across multiple datasets under varying sparsity levels.