Um estudo sobre a extração de características em dados do Twitter na tarefa de detecção de depressão

Abstract

Depression is a mental condition that affects millions of people worldwide, manifesting through persistent feelings of sadness, lack of interest, and changes in thought and behavior patterns. With the rise of social media, it has become possible to identify signs of depression through users' posts, offering new opportunities for the study of linguistic indicators associated with this condition. This study explores methods for feature extraction in social media data, aiming to identify signs of depression in Twitter users. The research began with the creation of a database from public posts, followed by the application of data preprocessing techniques. Cognitive Behavioral Theory was integrated into the theoretical framework, providing the basis for manual feature extraction. The selection of the most relevant features was carried out through hypothesis testing combined with the AdaBoost classifier. Among the key indicators found, the frequent use of first-person words and an increase in posts during nighttime by individuals labeled as depressed, compared to the control group, stood out. Additionally, data analysis revealed a lower number of effective words in the vocabulary of those labeled with depression. The results of this study have the potential to contribute to society. Health professionals can use data-driven screening tools and guide the creation of public mental health policies. Moreover, these techniques can assist social media platforms in identifying and supporting users at risk.