Sentiment and Emotion Classification Model Using Hybrid Textual and Numerical Features: A Case Study of Mental Health Counseling

Abstract

Mental health issues among individuals, particularly in counseling contexts, require practical tools to understand and address emotional states. This study explores the application of machine learning models for emotion detection in mental health counseling conversations, focusing on four algorithms: Bernoulli Naive Bayes, Decision Tree, Logistic Regression, and Random Forest. The dataset, derived from transcribed counseling sessions, underwent preprocessing, including stemming, stopword removal, and TF-IDF vectorization to create structured inputs for classification. Emotional categories such as "Depresi" (Depression), "Kecewa" (Dissapointed), "Senang" (Happy), "Bingung" (Confused) and "Stres" (Stress) were analyzed to evaluate model performance. Results indicated that Logistic Regression achieved the highest accuracy at 82%, showcasing its reliability and scalability, followed closely by Random Forest with 81%, demonstrating robustness in handling complex data structures. Bernoulli Naive Bayes performed competitively at 80%, excelling in computational efficiency, while Decision Tree recorded the lowest accuracy at 70%, reflecting its limitations in managing overlapping features and high-dimensional data. These findings highlight the potential of machine learning in addressing the increasing demand for scalable mental health support tools. The study underscores the importance of model selection, balanced datasets, and feature engineering to improve classification accuracy. Future work includes developing AI-driven chatbots for real-time emotion detection and integrating multimodal data to enhance interpretability. This research contributes to advancing automated solutions for mental health care, offering new pathways for timely and personalized interventions.