Ali Usman Abdullahi

A HYBRIDIZED BERT-BIDIRECTIONAL GRU FRAMEWORK FOR MITIGATING OVERFITTING AND ENHANCING GENERALIZABILITY IN E-COMMERCE SUPPLYCHAIN DEMAND PREDICTION

Aminu Adamu Ahmed¹; Ali Usman Abdullahi²; Abdulsalam Yau Gital¹; Abubakar Yusuf Dutse¹

Deep learning models applied to e-commerce demand prediction routinely achieve excellent training-set performance yet generalize poorly to unseen data - a failure mode known as overfitting that undermines their operational reliability. This study develops and evaluates a hybridized stacking framework that combines Bidirectional Encoder Representations from Transformers (BERT) for context-aware sentiment extraction with a Bidirectional Gated Recurrent Unit (Bi-GRU) for sequential demand forecasting, regularized through dropout, L2 weight decay, mini-batch training, and 5-fold cross-validation. The framework was trained and evaluated on two datasets: a primary survey-derived dataset (n = 114) and a large-scale secondary Kaggle e-commerce dataset (n = 500,000; test n = 75,000). Across both datasets, the hybridized Bi-GRU consistently outperformed standalone GRU and standalone BERT configurations. On the Kaggle dataset, the Bi-GRU achieved 98.86% sentiment-classification accuracy, a Mean Absolute Percentage Error (MAPE) of 2.78% for demand forecasting, and a train-validation loss gap of only 0.0049 - well below the 0.02 overfitting threshold applied in this study - compared with a gap of 0.2012 for standalone BERT on primary data. A systematic dropout-rate sweep identified 0.3 as the empirically optimal regularization rate on both datasets. Benchmarked against two recent comparable studies, the proposed framework reduced MAPE by 16.32 and 24.42 percentage points respectively. Feature-importance analysis further showed that behavioral constructs collectively accounted for approximately 68% of the Bi-GRU's predictive weight, evidencing a substantive theory-to-computation convergence. The findings offer both a validated architecture for overfitting-resilient demand prediction and a replicable regularization protocol for hybrid transformer-recurrent models operating on data-constrained e-commerce environments.