Depression is a global issue that has impacted millions of people worldwide. With an estimated over 264 million individuals across the globe suffering from depression, this condition is recognized as one of the most preva‎lent mental disorders. Depression affects people of all ages, genders, and backgrounds, serving as a significant cause of illness and disability in numerous cases and significantly impacting an individual's quality of life. The diagnosis of depression on social media has become an intriguing topic in the field of mental health research. With the increasing use of social media, researchers have become interested in using data from these platforms for efficient and cost-effective depression detection. Several studies have shown that the language used in social media posts can indicate an individual's mental state and the likelihood of experiencing depression. Machine learning algorithms and natural language processing techniques have been developed to analyze social media posts and identify patterns that could indicate depression. However, this approach comes with limitations, including privacy concerns and accuracy issues. Developing depression detection on social media has the potential to improve early diagnosis and intervention for individuals struggling with mental health issues. In essence, the goal of this research is to develop a novel method using machine learning and deep learning approaches for the purpose of detecting and predicting depression in users within the context of social media. This model assists in identifying a depressed user, even if the user hasn't directly mentioned being depressed in their posts. The model leverages the user's posted tweets within a specific time frame, the interaction pattern on social media (Twitter), observed emotional behaviors in tweets, and unique engineered concepts to predict user depression. The extracted features for classifying depressed and non-depressed users work well and provide an accuracy of 80% to 90%, which is a very good accuracy for the classification task. Additionally, various language models such as BERT have been applied to the tweet text. A hierarchical attention model has also been implemented. Finally, the outputs of BERT were combined with the outputs obtained from machine learning methods and the extracted features after several neural network layers. Ultimately, the achieved highest accuracy and precision metrics were 88.0% and 87.0%, respectively.

ناصر قديري مدرس

اعظم نقوي

عليرضا بصيري , مازيار پالهنگ