Leveraging Large-Language Models based Machine Learning for Sentiment Analysis and Regional Consumer Insights in Amazon Product Reviews

Abstract

In this paper, we use advanced NLP techniques and GPT based model, to analyze Amazon product reviews and gain actionable insights. Amongst the models chosen, DistilBERT was used as the feature extractor, BiLSTM for sentiment analysis and XGBoost for regional trend prediction was used to develop and evaluate a hybrid model. With an accuracy of 92.3% for positive insights and 87.6% for negative insights, the model proves to be reliable in understanding consumer sentiment. The F1 scores, particularly 90.5% for positive and 84.9% for negative insights, highlight its balance between precision and recall, ensuring that most relevant instances are correctly classified. Additionally, the low Mean Absolute Error (0.077 for positive and 0.124 for negative) further validates the model's capability to minimize prediction errors. Further, as a case study, we identified regional variations in customer sentiment for a portable Bluetooth speaker. The region-specific trend analysis reveals that the East Coast demonstrates the highest preference for the product, with approximately 70% positive reviews, indicating strong customer satisfaction in this region. On the other hand, the South Region exhibits the lowest positive reviews (~18%) and the highest proportion of negative reviews, highlighting significant dissatisfaction. The sentiment analysis from the analysis and predictions of the regional trend translates into marketing strategies and product improvement, especially in southern region to focus on sound quality and in rural markets where connectivity and battery performance need to be addressed. This approach demonstrates the promise of combining state of the art NLP techniques along with GPT to better understand customer preferences and support product development and marketing decisions.