Leveraging Unsupervised LDA and Supervised XGBoost for Advanced Restaurant Review Analysis

The proliferation of mobile technologies and internet accessibility has led to a surge in user-generated content, necessitating advanced methods for sentiment analysis in various domains, including restaurant reviews. This study proposes a novel approach combining Latent Dirichlet Allocation (LDA)-based topic modeling with Term Frequency-Inverse Document Frequency (TF-IDF) vectorization and an XGBoost classifier to enhance sentiment analysis performance. We analyze textual data by examining unigrams, bigrams, and trigrams to identify prevalent linguistic patterns and sentiment trends. The proposed methodology's efficacy is validated against a baseline corpus approach, with results demonstrating significant improvements across key performance metrics. The LDA-based TF-IDF approach, coupled with the XGBoost classifier, achieved an accuracy of 96.07%, a sensitivity of 95.43%, a specificity of 97.12%, and an F1-score of 96.16%, outperforming the corpus-based method in all aspects. These findings underscore the potential of integrating topic modeling with advanced feature extraction techniques and robust classifiers to advance sentiment analysis capabilities, offering reliable solutions for analyzing user feedback on digital platforms.