EST: A Semi-supervised ensemble approach for Explainable AI-driven stress detection in social media

Abstract

Social networking websites have become a vast ‘archive’ of human expression, containing hu man emotions from across the globe. Recognizing stress signals within this dynamic landscape is crucial for mental health monitoring, offering opportunities for timely intervention and sup port. This paper introduces a methodology integrating advanced machine learning techniques, notably Ensemble Learning and Self-Training, with Explainable Artificial Intelligence (XAI) to enhance stress detection capabilities. Our research follows a methodical approach, begin ning with foundational steps of data collection and exploratory data analysis, providing insight into sentiments within social media interactions. Prioritizing the interpretability of model pre dictions, our methodology aims to build trust and offer meaningful insights for end-users and mental health professionals. Subsequent phases involve data preprocessing, and refining textual data to extract subtle indicators of stress. The core of our methodology is the fusion of En semble Learning and Self-Training, strategically combining diverse learners to iteratively refine the model using labeled and unlabeled data. With the help of XAI, we seek to deliver clearly and straightforwardly noticeable insights relating to decision-making produced by our stress de tection model. Summing up the presented approach significantly contributes to mental health analytics taking a step forward toward better stress detection by a technical framework with an enhanced accuracy of 91.67%.