This research examines the efficacy and scalability of machine learning (ML) and deep learning (DL) models for sentiment analysis in public transit feedback. This study utilises a multilingual and unbalanced dataset from platforms like Google Reviews, Pantip, and Twitter to assess the merits and shortcomings of machine learning models, such as Logistic Regression and Random Forest, as well as deep learning models like CNN, BiLSTM, and fine-tuned BERT. Performance indicators, including accuracy, F1-score, and resource utilisation, underscore the trade-offs between computational efficiency and predictive precision. The results indicate that deep learning models surpass machine learning models in accuracy, with the fine-tuned BERT attaining the maximum accuracy of 90.45%, although with increased resource usage. In contrast, ML models are more appropriate for resource-limited settings, providing rapid training and inference with acceptable accuracy. The research emphasises the need of matching model selection with deployment contexts and suggests hybrid methodologies to enhance performance and scalability.