Depression Detection on Social Media X by Text Analysis with Attention-based CNN-BiLSTM using FastText

The exponential growth of the digital domain has rendered contemporary society reliant on social media. Conse- quently, the manner in which many individuals engage with social media can manifest indications of distress, such as depression. Social media X is a popular platform that can contain all the outpourings of its users called tweets. With the increasing cases of depression, it is important to be able to detect depression early. This research contributes to combining a hybrid deep learning method to detect depression on social media X with TF-IDF as a feature extraction that plays a role in measuring the importance of words in each user’s tweet, FastText as feature expansion to improve word representation and finding semantic similarities, and attention mechanisms as optimization in adding weights. With a total of 50,523 tweet data, a similarity corpus of 100,594 was constructed. Based on the result, using the attention mechanism the BiLSTM model achieved 84.25% accuracy, a 2.03% increase from the baseline, and the CNN-BiLSTM hybrid model achieved 83.01% with an accuracy increase of 0.99% from the baseline.