A novel hybrid clustering and classification framework for multi-level depression detection in social media texts

Depression is a prevalent psychiatric condition worldwide, with significant social and economic implications. Despite its high incidence, many individuals with depression remain undiagnosed and untreated. Meanwhile, people increasingly use social media platforms to express their emotions and thoughts. Consequently, leveraging these platforms for depression detection may help address several related challenges. This paper proposes a three-stage methodology, based on text mining techniques, to determine the severity of depression in individuals who post textual content on social media. In the proposed framework, each post is transformed into a vector of numerical features using established feature extraction methods. Principal Component Analysis is then applied to select the most informative features for identifying whether a post indicates depression via a classification algorithm. If depression is detected, the method clusters the relevant posts based on their characteristics, grouping similar texts together. A second classification model is then applied within each cluster to determine the level of depression. To equip the model with effective algorithms at each stage, the Taguchi method is used to identify the best combination of feature extraction, clustering, and classification techniques. Specifically, Bidirectional Encoder Representations from Transformers (BERT) is used for deep contextual feature extraction, Deep Embedded Clustering (DEC) is employed for clustering, and Support Vector Machine (SVM) is used for classification. Numerical results show that the proposed approach can accurately classify individuals’ posts into one of four depression levels: non-depressed, low, moderate, and severe. These findings suggest that social networks offer a platform for assessing mental health through textual analysis.