<p>Predicting user preferences in sequential recommendation systems is challenging due to complex and diverse periodic behaviors across user groups. Most existing models operate in the time domain and often miss periodic regularities, especially when periodicity differs across users. To overcome these limitations, we propose TFRec: a Time-Frequency Model for capturing periodic preferences in sequential recommendation. Specifically, the time-domain branch uses a multi-head attention backbone augmented with Two-Stage Feature Modulation (TFM), which alleviates single-peak attention and yields balanced interaction representations. The frequency domain branch employs a Dual-Kernel Periodic Encoder (PE) that adaptively selects either a Shared Periodic Kernel (SPK) or an Individual Periodic Kernel (IPK) according to a learned periodicity score, capturing cohort-common and user-specific patterns, respectively. This is followed by an improved Squeeze-and-Excitation mechanism that dynamically recalibrates spectral features under diverse periodic patterns. The two branches are fused to form the final representation. Extensive experiments on six public datasets show that TFRec outperforms eight state-of-the-art baselines, demonstrating its effectiveness in capturing dynamic and periodic user preferences. To accommodate the cost of the frequency domain transforms and periodic kernels, we implement TFRec with GPU-accelerated HPC, which enables scalable training and inference and supports responsive recommendation settings.</p>