Sentiment Analysis for Mental Health Using Boosting, Bagging, and DeBERTa from Social Media Data

Nowadays, social media influences from misinformation to cyberbullying and from influencing anxiety and depression to suicidal tendencies have been quite high. All such conditions require proper and timely identification for suitable intervention. The classes used for this work are “Anxiety”, “Bipolar”, “Depression”, “Normal”, “Personality Disorder”, “Stress”, and “Suicidal”. The text was cleaned, under-sampling was done to make the class labels balanced, and labels were encoded. The research employs machine learning models such as Random Forest, Gradient Boosting, LightGBM, and DeBERTa, using an 80-20% train-test split and undersampling to balance class labels. The objective of this research is to assess the effectiveness of boosting, bagging, and the DeBERTa model in classifying mental health conditions based on social media data. The study achieved very good performance using the DeBERTa model: an accuracy of 0.8741, precision of 0.8749, a recall of 0.8741, and an F1-score of 0.8740. However, classes such as “Depression” versus “Suicidal” were slightly overrated. It is a take-home message from the current study that machine learning might effectively enhance diagnostic-related to mental health-more precisely by the DeBERTa algorithm. This study advances the understanding of classifying mental health conditions with a focus on improving detection between closely related conditions like depression and suicidal tendencies.