A Threshold-Optimized Hybrid Ensemble for imbalanced Aspect-Based Sentiment Analysis of Restaurant Reviews
TL;DR
A hybrid lexical-probabilistic ensemble system that combines SentiWordNet lexical resources with a dual-branch Naïve Bayes ensemble, using Multinomial Naïve Bayes for text features and Gaussian Naïve Bayes for continuous lexical sentiment features, to enhance accuracy across five restaurant aspects.
Abstract
Aspect-Based Sentiment Analysis (ABSA) is a vital method for extracting detailed opinions from customer feedback, but current approaches often overlook important factors such as menu variety and struggle with class imbalance in real-world data. This research propose a hybrid lexical-probabilistic ensemble system that combines SentiWordNet lexical resources with a dual-branch Naïve Bayes ensemble, using Multinomial Naïve Bayes for text features and Gaussian Naïve Bayes for continuous lexical sentiment features, to enhance accuracy across five restaurant aspects: food quality, service, physical environment, price fairness, and menu variety. The study was evaluated on 10,000 Kaggle restaurant reviews. After excluding neutral reviews, the remaining binary dataset was imbalanced, and evaluation was performed using a stratified 80:20 train–test split. Dataset preprocessing through tokenization, lemmatization, and stop-word removal was applied, and aspect extraction was also performed using domain-specific keyword dictionaries enriched with WordNet synonyms, Part-of-Speech tagging, and dependency parsing, while SentiWordNet assigned aspect-level sentiment scores. A threshold optimization strategy was applied to the posterior probabilities of the positive class to improve minority-class recall while maintaining overall precision without harming majority class performance. Results show that the ensemble achieves 86% precision, 80% recall, and 82% F1-score, significantly outperforming basic classifiers such as standalone Naïve Bayes, SentiWordNet-augmented Naïve Bayes, and a hybrid ELMo-Wikipedia method. Statistical tests with the Wilcoxon signed-rank test confirm these improvements (p < 1.91e-06). This framework provides fine-grained aspect-level sentiment analytics for restaurant decision support and lays a solid groundwork for handling imbalanced sentiment analysis. Future work discusses integrating transformer-based models and multilingual support. Keywords: Aspect-Based Sentiment Analysis, Class Imbalance, Customer Satisfaction, Ensemble Learning, Naïve Bayes Classifier, Restaurant Reviews, SentiWordNet, Threshold Optimization