Detecting Emotional State of Depression in Social Media Posts Using Logistic Regression-Recursive Feature Elimination

With the rise of global mental health issues and the growing use of social media, detecting depression through social media text data has become increasingly crucial. Although many studies have focused on identifying depression by analyzing social media texts, most of these studies mainly focus on a single feature extraction technique. However, the effective combination of multiple feature extraction techniques for depression detection remains underexplored. This study aims to explore how various feature extraction methods can be effectively combined to enhance depression detection and to develop an optimized feature selection method that improves model performance without losing essential information. Advanced feature extraction methods—TFIDF, N-gram, SentiWordNet, and BERT—were combined with machine learning algorithms to analyze social media texts using the eRisk dataset. Three experiments were conductedthe first evaluated individual feature extraction methods across models such as Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM); the second assessed the effectiveness of combining multiple features; and the third applied Logistic Regression-Recursive Feature Elimination (LR-RFE) to optimize feature sets. The results indicated that combining TFIDF, N-gram, SentiWordNet, and BERT significantly enhanced performance, particularly in the LR model. Additionally, applying LR-RFE further enhanced model accuracy and reliability across all tested machine learning modelsachieved an accuracy of 0.800 and an F1 score of 0.7969, which demonstrating improved generalizability and robustness. These findings underscore the importance of comprehensive feature extraction and feature selection for effective depression detection through social media text data.