In the digital age, recommendation systems play a vital role in alleviating information overload, enhancing user engagement, and driving significant growth in e-commerce. While widely adopted in next-basket prediction tasks, conventional methods primarily focus on short-term user interactions and often overlook long-term behavioral patterns that are crucial for delivering personalized recommendations. To address this limitation, we propose a novel approach: Personalized Item Frequency (PIF), a key feature that models users’ repeated purchase behaviors over time. Integrating PIF allows for the capture of subtle and consistent buying habits, thereby improving recommendation accuracy beyond traditional frequency-based or recency-oriented models. Building on this foundation, we introduce PIFTA4Rec, a hybrid neural network model designed to enhance both recommendation precision and computational efficiency. PIFTA4Rec combines K-Nearest Neighbor (KNN) for PIF-based vector prediction with a Temporal Attention mechanism to accurately model the timing of user purchases. In addition, the model leverages the multi-head attention mechanism of Transformers to capture complex and dynamic user–item relationships across diverse contexts, making it particularly effective for next-basket recommendation scenarios. Empirical evaluations on two real-world datasets demonstrate that PIFTA4Rec consistently outperforms state-of-the-art next-basket recommendation models in terms of both accuracy and robustness. These results underscore the importance of incorporating long-term purchase patterns—such as those captured by PIF—in advancing next-basket recommendation systems. This study introduces a unified and interpretable next-basket recommendation framework by integrating traditional PIF-based modeling with deep learning techniques, delivering both theoretical insights and practical benefits for future research and applications.
