Comparative Analysis of Depression Comment Classification Using Traditional Machine Learning and Transformer-Based Models with Easy Data Augmentation

Depressive disorders are a pressing global mental health issue, affecting over 280 million people and requiring early intervention. In the digital age, social media text analysis has become a valuable tool for early symptom detection due to increasing online expressions of depression. However, analyzing social media text is challenging due to inherent characteristics such as informal language, slang, and extensive Indonesian and English code-mixing. This study performs accurate binary classification of social media comments (Depression vs. Normal) on a large-scale multilingual dataset featuring Indonesian and English code-mixing. We rigorously compare four models Multinomial Naïve Bayes (MNB) and Support Vector Machine (SVM) as classic baselines, and IndoBERT and XLM-RoBERTa as contemporary Transformer models. Crucially, all models were tested in two scenarios No-EDA and With-EDA (Easy Data Augmentation). The results show that the XLM-RoBERTa model without augmentation achieved superior performance with an F1-Score of 97.6% and an accuracy of 97.4%. This highlights XLM-RoBERTa's strength in handling cross-lingual complexity. Conversely, simple lexical augmentation (EDA) proved ineffective and caused performance decline in monolingual Transformer models (IndoBERT), indicating limitations in simple lexical augmentation for complex languages. Overall, this research provides strong recommendations for developing an optimal AI-based early depression detection system adaptive to the multilingual (Indonesian-English) context and linguistic diversity of Indonesia.