Identifying Depression Based on Posting Pattern on Social Media Using the ULMFiT Method

This research focuses on the application of ULM-FiT for a classification task of depressed and normal X users based on their tweets. A three-layer AWD-LSTM architecture is trained using 109.112 Indonesian Wikipedia pages to initiate the language model layer, then trained with 232.523 lines of tweets for fine-tuning, followed by training a binary classification layer with 79.069 rows of tweets from 117 labeled users, and preset learning rates were used to support in detecting depressed users. The result of the model produced a general language model with a perplexity of 32.57. The best classification model achieved an accuracy of 57%, 59% precision, and an F1 score of 56%. This study attempts to detect Depression in X users through textual information from social media posts using the ULMFiT method. Future work should explore the subtle, context-dependent, and cultural linguistic patterns of mental health issues expressed in social media while also applying psychological theories to detect depressed users in social media.