نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسنده English
Accurate daily demand forecasting at the customer-product level remains a critical yet challenging task in retail supply chain management. This challenge is mainly driven by sparse transaction records, highly volatile consumer behavior, and the limited use of unstructured product information in forecasting models. To address these issues, this study proposes a novel multimodal semantic Transformer framework that integrates semantic product embeddings, extracted from product descriptions through natural language processing, with structured customer purchase history. In addition, the model incorporates multi-scale temporal features to capture seasonality, short-term demand fluctuations, and longer-term purchasing patterns more effectively. The proposed framework is evaluated on a real-world online retail dataset and compared with several benchmark models, including Long Short-Term Memory (LSTM) networks, Gradient Boosting Machines, and a unimodal Transformer architecture. The empirical results show that the proposed model consistently outperforms all baseline methods and achieves a 15.5% reduction in Mean Squared Error (MSE). The findings further reveal that semantic product information provides valuable predictive signals, particularly for low-velocity products where historical demand patterns alone are often insufficient. Moreover, temporal embeddings combined with dynamic attention conditioning improve the model’s ability to represent seasonal variations and evolving customer preferences over time. Overall, the results confirm that deep multimodal integration offers an effective, scalable, and interpretable solution for fine-grained retail demand forecasting and can support more customized inventory decisions and more resilient, customer-centric retail supply chains.
کلیدواژهها English
