Predicting Customer Satisfaction Through Layered NLP Techniques: Traditional Meets Deep Learning

Abstract

This study presents a robust hybrid artificial intelligence framework designed to predict customer satisfaction based on online reviews a critical capability for enhancing business performance and optimizing customer relationship management (CRM) practices. We propose a unified architecture with high precision for sentiment classification, utilizing the "All Beauty (2023)" dataset from Amazon Reviews. The approach synergistically combines conventional machine learning (ML) algorithms with state-of-the-art deep learning (DL) techniques. Although standalone ML and DL models have demonstrated promising outcomes in sentiment analysis, they often fall short in addressing intrinsic data challenges such as class imbalance and linguistic subtleties. To overcome these limitations, we introduce HybridStack, a novel ensemble architecture that integrates diverse ML and DL models to significantly improve classification accuracy. The methodology encompasses rigorous text preprocessing, a dual-path feature extraction pipeline (TF-IDF for ML models, and embedding techniques for DL models), and a meta-classifier to consolidate predictions into three categories: dissatisfied, neutral, and satisfied. To address class imbalance, we employ weighted loss functions alongside oversampling strategies. The proposed HybridStack model achieved an outstanding and unparalleled F1-score of 0.6889 and an accuracy of 0.8593, substantially outperforming all baseline individual models. This hybrid approach highlights the benefits of algorithmic synergy and demonstrates its effectiveness in generating more granular and actionable insights, thereby supporting the development of more refined CRM strategies and fostering stronger customer relationships.