<p>Sequence recommendation (SR), an important research direction in personalized recommender systems, faces a threefold challenge: sequence noise, users’ dynamic and periodic (seasonal) preferences, and feature over-smoothing. To address sequence noise, numerous denoising methods have been explored; however, most existing studies do not simultaneously consider the dynamic changes in users’ interests and their periodic patterns. In this paper, we introduce a novel SR model, Adaptive Sequential Recommendation with Dynamic and Periodic Preferences Capturing (ASRec), which performs implicit denoising to reduce noise, captures users’ dynamic and periodic preferences, and alleviates feature over-smoothing in an efficient manner. This approach aims to more accurately reveal users’ true interests and improve recommendation performance. In our model, we adapt the denoising strength to dataset characteristics via a Fourier Smoothing Module (FSM) to minimize noise interference. Meanwhile, we deploy a Hybrid Attention Mechanism Module (HAMM) that builds upon and improves the low-pass filtering property of self-attention, enabling the joint capture of users’ dynamic and periodic preferences. In addition, it reduces feature over-smoothing, thereby better modeling users’ true preferences. We conduct extensive experiments on five real-world datasets and show that ASRec significantly outperforms state-of-the-art approaches in both effectiveness and efficiency. To facilitate reproducibility, we publicly release the code and datasets for ASRec at <a href="https://github.com/WLHao0313/ASRec">https://github.com/WLHao0313/ASRec</a>.</p>