Exploring Hybrid Approaches for Sentiment Classification: A Comparative Study of LSTM, Naive Bayes, and Bayesian Network on IMDB Reviews

This paper evaluates the use of different types of models for machine learning and deep learning applied to sentiment analysis on the IMDb movie reviews dataset. Classical approaches, including Naive Bayes, Logistic Regression, and XGBoost, are compared with sequential deep architectures like (LSTM) and Recurrent Neural Networks (RNN). A Bayesian network is added to the portfolio in order to provide probabilistic inference for sentiment prediction. Techniques applied to preprocess the data include BoW and tokenization with padding. Results show that LSTM and RNN significantly identify sequential text patterns and perform well at high accuracy and F1 scores especially if input sequences are tokenized and padded. Classical models like Naive Bayes and Logistic Regression achieved competitive precision and recall capabilities with good computational efficiency and interpretability by using BoW features. XGBoost proved to be robust with optimized gradient-boosting techniques, while the Bayesian network presented a novel probabilistic approach, enabling the system to provide a probability of sentiment prediction based on the relationship between variables of sentiment and prediction. Results bring to light the potential integration of deep models with the capabilities of sequential learning and classical approaches or probabilistic models such as Bayesian networks for developing a scalable and efficient sentiment analysis framework for real-world applications. The experimental setup involves using LSTM, XGBoost, Naive Bayes, Logistic Regression and Bayesian Networks models to analyze IMDB movie reviews for sentiment classification has achieved 85.25, 86.13, 83, 90.21 and 80 accuracies respectively.