Depression Detection on Indonesian Social Media Using Fine-Tuned IndoBERT and SVM
TL;DR
Given statistically equivalent accuracy, combining fine-tuned IndoBERT embeddings with SVM offers substantially lower storage requirements at no measurable cost in classification performance, making it a promising, resource-efficient approach for depression detection on Indonesian social media.
Abstract
Depression has become a major mental health issue in Indonesia, where approximately 167 million of the country’s 273 million citizens actively use social media platforms such as X (Twitter). The informal writing style, code-mixing, and linguistic variability in Indonesian tweets create significant challenges for automated depression detection systems. This study evaluates a fine-tuned IndoBERT model combined with a Support Vector Machine (SVM) classifier for detecting depression-related indications from Indonesian-language tweets. A total of 10,082 Indonesian tweets were collected and labeled into two categories: Terindikasi Depresi and Tidak Terindikasi; after deduplication, 3,874 unique tweets were used for modeling. Two scenarios were compared: (1) a fine-tuned IndoBERT model, and (2) fine-tuned IndoBERT CLS embeddings with a linear SVM classifier. The fine-tuned IndoBERT model achieved 73.20% accuracy (AUC-ROC = 0.8212), while the hybrid approach achieved a marginally higher 73.71% accuracy (AUC-ROC = 0.8088); a McNemar’s test found this difference not statistically significant (p = 0.86). Both models outperformed five traditional TF-IDF-based baselines (best: 70.36%) on the same held-out test set. The hybrid model required only 0.01 MB of storage versus 475.24 MB for the full fine-tuned model. Given statistically equivalent accuracy, combining fine-tuned IndoBERT embeddings with SVM offers substantially lower storage requirements at no measurable cost in classification performance, making it a promising, resource-efficient approach for depression detection on Indonesian social media.