Accurate click-through rate (CTR) prediction is essential for recommender systems. However, three inherent challenges constrain current methodologies. First, most methods learn fixed feature representations, overlooking feature importance variations across instances. Second, while self-attention mechanisms can effectively capture cross-feature relationships, their indiscriminate treatment of all feature pairs may introduce noise into the learning process. Third, although many methods automatically learn higher-order feature interactions, noise inevitably accumulates with increasing interaction orders. To address these challenges, we propose the AFRNS model (Adaptive Feature Refinement and Noise Suppression), which integrates two key components: the Adaptive Feature Refinement Module (AFRM) and the Dual-Stream Gated Factorized Interaction Network (DS-GFIN). AFRM adaptively refines feature representations based on instance-specific contextual information, with Multi-Head Adaptive Sparse Self-Attention (MHASSA) as one of its core components to effectively suppress noise. DS-GFIN captures high-order feature interactions and designs an information gate to dynamically filter important interactions, thereby suppressing noise. Importantly, AFRM’s modular design allows seamless integration into CTR models, improving their accuracy. Extensive experiments conducted on three benchmark datasets demonstrate the superior effectiveness of AFRNS.