Adaptive Deep Learning Framework for Fair and Scalable E-Commerce Customer Segmentation in Data-Scarce Environments

Abstract

In e-commerce, where new procedures come every day, customer segmentation provides direct avenues for personalized marketing and strategic decision-making. However, traditional segmentation techniques face major challenges in data-scarce environments, especially in remote regions where purchasing behaviours and customer characteristics vary vastly. This study proposes a new ensemble deep learning framework that introduces LSTM networks for sequential purchase analysis and utilizes NLP for sentiment-based customer insights. The model uses transfer learning for adaptation on different datasets, whereas Bayesian deep learning helps quantify uncertainty and, in turn, is about enhancing the robustness of the model. Some fairness-aware mechanisms ensure that no group is unfairly burdened with segmentation; the online learning module continuously updates the model with real-time data streams during model execution. Experimental greatness on real-world e-commerce datasets shows that our approach outperformed traditional systems, achieving an average accuracy of 94.7%, precision of 93.5%, recall of 92.3%, and an F1-score of 92.8%, surpassing classical Random Forest, XGBoost, and standard LSTM models by huge margins. The proposed algorithm represents a fair and scalable strategy to enhance data-driven marketing decisions in tackling potential customers, particularly in the underrepresented segment of the market.