Sentiment Analysis of Student Perceptions of Generative AI using Data Augmentation and Machine Learning Models

Abstract

The development of Generative AI has significantly changed how students access information, understand learning materials, and generate ideas in higher education. Although Generative AI supports independent learning, it also raises concerns regarding overdependence, declining critical thinking skills, and academic integrity. This study evaluates the performance of sentiment analysis models using a processing pipeline that incorporates lexical, semantic, and generative data augmentation. The main challenge addressed in this study is class imbalance, particularly the limited number of negative sentiment samples compared to positive and neutral classes. This study applies an experimental quantitative approach consisting of dataset preparation, text preprocessing, data augmentation, feature extraction using TF-IDF, model training, and evaluation using Stratified K-Fold Cross Validation. The machine learning models evaluated include Multinomial Naive Bayes, Logistic Regression, Random Forest, and Linear Support Vector Machine. The experimental results show that Linear SVM achieved the best performance, with an average accuracy of 79.07% and a weighted F1-score of 73.90%. Compared descriptively with the non-augmented baseline, Linear SVM showed an observed increase in accuracy from 65.00% to 79.07% and in weighted F1-score from 63.03% to 73.90%. Data augmentation also enabled partial recognition of minority-class sentiment, although a substantial proportion of negative and positive samples were still misclassified as neutral. These findings indicate that hybrid data augmentation can support the performance of classical machine learning models on small and imbalanced educational text datasets, particularly when combined with TF-IDF and Linear SVM. However, a post-hoc audit identified a discrepancy between the class distribution of the original dataset and that of the final processed dataset. Therefore, the observed model performance should be interpreted as the result of the overall processing pipeline rather than as the isolated effect of data augmentation.