Enhancing E-Commerce Sentiment Analysis: An Ensemble Approach To Addressing Contextual Data Imbalance

Sentiment analysis plays a vital role in managing product quality in e-commerce. Customers rely on reviews to gauge others’ opinions about products. However, existing sentiment analysis models face several challenges. These include difficulties in capturing the complexity and context of language, leading to reduced accuracy. The imbalanced datasetswhere one sentiment class dominates-pose a challenge, often resulting in biased classification. This paper proposes a hybrid model combining BERT, SMOTE, and VADER to address these challenges. BERT provides a deep understanding of local and global dependencies in text, SMOTE generates synthetic data to address class imbalance, and VADER provides labels for fine tuning BERT. As ensemble they improve sentiment classification accuracy and reduce bias. We conducted experiments on a realtime dataset of e-commerce product reviews, covering various product categories and sentiment distributions. The proposed model achieved $99.19 \%$ accuracy outperforming traditional machine learning models along with high F1 scores, precision, and recall metrics.