Prediction of Customer Sentiment Based on Online Reviews Using Machine Learning Algorithms

Abstract

Customer opinions and feedback play a pivotal role in enhancing business operations and decision-making processes. Sentiment analysis is a crucial technique used to decipher customer opinions from their feedback and thus provide valuable insights for businesses. However, analysing and understanding reviews is an intricate process and prone to be misleading if not conducted meticulously. This study aims to extract and classify customer emotions from e-commerce reviews of women’s clothing in terms of polarity of sentiment, enhancing sentiment analysis accuracy by means of machine learning (ML) classifiers. In addition, the study addresses the challenge of imbalanced data samples. Several supervised ML models, including Naïve Bayes, Random Forest, Support Vector Classifier (SVC) and Extreme Gradient Boosting (XGB), were employed for sentiment analysis. The study also attempts to deal with negations, and pre-processing steps were implemented to reduce the dimensionality and noise of the raw text. In SVC and XGB models, negation handling significantly improved the precision and recall values of minority classes. Moreover, a hybrid class balancing approach and the synthetic minority oversampling technique (SMOTE) were adopted. The findings indicate that the XGB classifier, combined with SMOTE sampling, produced the most accurate results, yielding better F1 and ROC AUC scores.