Fine-Grained Attribute Decoupling and Interest Denoising Network for Click-Through Rate Prediction

Click-through rate (CTR) prediction plays a critical role in personalized recommendation systems. Accurate CTR prediction not only enhances user experience and satisfaction but also brings substantial commercial value to online service platforms such as e-commerce. A key challenge in CTR prediction is the precise representation of user interest preferences. Existing methods mostly focus on item-level user interest modeling, neglecting the complex relationships between fine-grained attributes (such as category, brand, etc.), making it difficult to effectively address the noise problem caused by item attribute coupling and user interest drifting. Furthermore, identifying and distinguishing hard negative samples from noisy ones remains a critical issue that needs to be resolved. To address these challenges, we propose a fine-grained attribute decoupling and interest denoising network. This network effectively mitigates noise caused by attribute coupling and interest drifting through joint modeling of fine-grained attribute decoupling and interest denoising. Specifically, the network decouples item attributes to model user interests with a fine-grained attribute-aware interest denoising module, which handles noise caused by attribute coupling and user interest drifting. To further optimize user interest representation, we design a contrastive interest optimization module based on hard sample enhancement, ensuring a more accurate and comprehensive user interest representation. We conduct experiments on three real-world datasets and compare the proposed method with baseline approaches, validating its effectiveness.