In real-world scenarios, user interests inherently exhibit dynamic and multifaceted characteristics. While Transformer-based sequential recommendation models have achieved remarkable success in modeling long-term behavioral dependencies through self-attention mechanisms, our analysis reveals two critical limitations: (i) The inherent long-term dependency focus tends to overshadow transient behavioral dynamics, particularly diminishing sensitivity to emerging short-term preferences; (ii) Recent spectral analyses demonstrate that standard Transformers possess intrinsic low-frequency bias, resulting in systematic neglect of high-frequency signals within feature representations. To tackle these two challenges, we introduce DISRec (Dual-Interest Sequential Recommender), a novel hybrid architecture that synergistically integrates global and local behavioral patterns. The framework preserves Transformer’s superior long-term interest modeling capabilities while introducing a parallel temporal convolutional pathway to capture immediate preference signals through localized feature extraction. Furthermore, we devise a spectral enhancement module employing Euler Fourier Transformations combined with Fast Fourier Transform (FFT) operations, enabling explicit learning of periodic patterns in both time and frequency domains. Thorough evaluations on three benchmark datasets highlight DISRec’snotable advancements over leading Transformer variants, especially in situations demanding quick adaptation to preference changes. Ablation studies further validate that our dual-interest modeling and spectral processing components contribute synergistically to performance gains.