Abstract:In the research on recommendation systems, data sparsity and user feature interaction have always been the difficulty. This study proposes a dual-tower model that integrates SENet-based feature recalibration and an attention mechanism. This model achieves the efficient candidate set recall via a dual-tower structure and utilizes complex feature interaction for precise ranking. Based on the conventional recommendation frameworks, the model introduces an improved attention mechanism to enhance dynamic interaction capabilities. Additionally, it integrates a SENet module to adaptively adjust feature importance, and employs a hybrid architecture that combines the advantages of matrix factorization and deep learning to further improve feature representation and generalization abilities. Experiments conducted on the MovieLens-1M and Netflix datasets demonstrate that the proposed model outperforms mainstream baseline models in both rating prediction and classification tasks, validating its superiority in extracting user features and mitigating data sparsity.