Abstract
Mental health sentiment classification from textual data has attracted increasing attention as a computational approach to support large-scale psychological assessment; however, multi-class classification remains challenging due to noisy text, class imbalance, and semantic overlap among categories. This study proposes and evaluates a machine learning framework for seven-class mental health sentiment classification that integrates enhanced text preprocessing with lemmatization, data augmentation via back-translation, TF-IDF feature extraction, and systematic model evaluation across multiple classifiers, including Logistic Regression, Decision Tree, Random Forest, Gradient Boosting, K-Nearest Neighbors, AdaBoost, and XGBoost, under three hyperparameter tuning strategies: Grid Search, Random Search, and Particle Swarm Optimization (PSO). Experimental results indicate that ensemble-based models consistently outperform single classifiers, with the PSO-optimized Random Forest achieving the best numerical performance, attaining an accuracy of 0.933, a macro F1-score of 0.923, and a ROC AUC of 0.989, demonstrating strong generalization and balanced class-level performance despite dataset imbalance. These findings confirm that the combination of robust preprocessing and metaheuristic-based hyperparameter optimization significantly enhances multi-class mental health sentiment classification and supports its potential use as a scalable decision-support tool for large-scale mental health screening, while not intended to replace clinical diagnosis.