Supervised learning algorithms for subject- and textlevel classification of depression-related context among Thai-language Facebook users

Abstract

Background: Depression is an increasing concern worldwide that requires urgent attention, particularly in Thailand. Without timely intervention, it can lead to serious consequences, such as health complications, reduced social participation, and even suicide. Consequently, research has increasingly focused on developing algorithms to identify and classify social media data into depression severity.

Objectives: This study aimed to examine label mismatches between self-assessment forms and social media behaviors and to evaluate the feasibility of using experts’ opinions to classify Thai Facebook text into depression categories.

Methods: The methodology involved data collection through surveys and Facebook scraping, followed by data preprocessing, feature engineering, and model training and validation. The models employed included the support vector machine, logistic regression, and decision tree, applied at the subject and text levels. Labels were derived from the Thai Depression Inventory (TDI) and experts’ evaluations.

Results: Only 30 out of 60 participants actively used Facebook during the study period, with substantial variation in posting behavior. Subject-level classification using TDI labels yielded better, more balanced performance metrics than classifications based on experts’ opinions at the subject and text levels. Part-ofspeech tagging showed slight correlations with depression severity, whereas emotions did not.

Conclusion: Subject-level classification using self-assessment labels performed better than expert-derived labels and text-level classification when compared to a dummy baseline model. This suggests a slight correlation between the self-assessment forms and social media behaviors. In contrast, the poor agreement between labeling approaches suggests that text data alone may be insufficient as a reliable clinical indicator of depression.