DEVELOPMENT OF A SEMANTIC FEATURE EXTRACTION AND PSO-BASED FEATURE SELECTION MODEL WITH LSTM AND RNN CLASSIFICATION FOR DEPRESSION IN SOCIAL MEDIA TEXT

Abstract

Social media text is short, informal, and noisy, and these properties make the feature engineering stage difficult in depression-related text analysis. This study focuses on semantic feature extraction and feature selection. A GloVe feature extraction and PSO feature selection framework is developed in which Global Vectors for Word Representation (GloVe) converts preprocessed tweets into dense semantic vectors, and Particle Swarm Optimization (PSO) selects the most informative dimensions from that vector space. The optimized feature set is validated using two comparative configurations, namely GloVe-PSO with LSTM and GloVe-PSO with RNN, to verify whether the selected features improve class separability. Kaggle dataset used in this study contains more than 20,000 posts; after cleaning and filtering, 9,075 records were retained for embedding and validation. The original 100-dimensional GloVe representation was reduced to 43 selected semantic features. The optimized feature set achieved strong downstream validation performance, with RNN giving the best evaluation: 89.37% accuracy, 90.66% precision, 96.79% recall, and 93.62% F1-score, while the feature extraction-selection stage required 30 seconds. The GloVe-PSO with RNN configuration outperformed GloVe-PSO with LSTM in the comparative validation. The results indicate that optimizing semantic embedding dimensions can reduce redundancy and improve the practical usefulness of features for depression-related social media analysis.