Social networking sites have become widespread “archives” of human expression, reflecting emotions worldwide. Detecting stress signals within this dynamic environment is essential for monitoring mental health and providing timely support. This paper presents a methodology that combines advanced machine learning techniques, specifically Ensemble Learning and Self-training with Explainable Artificial Intelligence (XAI) to improve stress detection. The approach begins with data collection and exploratory data analysis to capture sentiment insights from social media interactions. Prioritizing interpretability, the methodology aims to provide actionable insights for end-users and mental health professionals. Subsequent stages focus on data preprocessing and refining textual data to identify subtle indicators of stress. At its core, the methodology integrates Ensemble Learning and Self-training, leveraging diverse models to iteratively refine predictions using both labeled and unlabeled data. With XAI, this research delivers transparent and easily understandable insights into the model’s decision-making. Overall, this approach advances mental health analytics by enhancing stress detection capabilities, achieving an improved accuracy of 91.67%.