<p>Personalized recommendation systems play a critical role in helping users discover relevant content amidst information overload. However, existing methods typically treat item attributes in a coarse-grained manner and suffer from entangled preference representations, which limits their ability to distinguish true user interests from noise and fails to capture global periodic patterns. This paper proposes AdaDisen, an enhanced sequential recommendation framework that addresses key limitations in existing approaches. By constructing a fine-grained type-aware heterogeneous graph that incorporates multidimensional contextual information, we first learn initial user/item representations using a heterogeneous graph attention network with independent relation channels. We then obtain user preferences from three distinct views by incorporating global frequency preferences (utilizing fast Fourier transform to capture long-term periodicities and mitigate noise) alongside orthogonality-regularized dynamic interest preferences (modeling temporal behavioral patterns) and static attribute preferences (capturing stable trait-based inclinations). Furthermore, an adaptive multi-view fusion module is introduced to automatically learn personalized importance weights to balance these components during prediction. Experiments on Amazon-Books, Amazon-Beauty, and MovieLens-1M datasets demonstrate that AdaDisen achieves better performance than representative baselines in HR@10 and NDCG@10 metrics. Our model reduces representation entanglement, enhances preference modeling granularity, and improves both recommendation accuracy and interpretability by better separating complementary preference factors underlying user decisions. Despite its multi-view design, AdaDisen remains computationally practical for sequential recommendation by integrating lightweight complementary modules within a unified framework.</p>