Click-through rate (CTR) prediction is crucial to recommendation systems. In recent years, numerous deep learning models have been effectively used for CTR prediction, aiming to effectively extract diverse user interests from user behavior. However, these models often face two problems: (1) Existing models are insufficient in capturing the dynamic changes in user interests, failing to simultaneously focus on users’ explicitly multiple interests and hidden latent interests. (2) Most models emphasize learning implicit feature interactions through deep neural networks while neglecting explicit feature interactions between each user interaction item and the candidate item. To tackle these problems, we design a new model, the Hybrid Dynamic Interest Extraction and Feature Co-action Network (DIFCN). First, we designed a Dynamic Interest Extraction Module comprising three different sub-networks. Specifically, we design the Behavior Refiner Network (BRN) for obtaining more accurate representations of user interests, and the Multi-Interest Extraction Network (MIN) and Hidden Interest Extraction Network (HIN) to dynamically extract explicit multiple interests and hidden interests of users. Then, we designed a Feature Co-action Module to model explicit feature interactions between user interaction items and candidate items. Our experiments confirm that our proposed DIFCN surpasses other models in the CTR prediction task.