Mental Health Symptom Analysis: A Combined Word2Vec, PCA, and K-Means Clustering Study of Indonesian Facebook Narratives

Abstract

Mental health problems are increasingly expressed through social media narratives, including Indonesian Facebook communities. However, previous clustering studies often focus on broad discourse themes, user-risk detection, or community-level patterns rather than symptom-oriented mental health expressions. This study proposes an unsupervised symptom-oriented clustering approach for Indonesian Facebook mental health narratives by integrating DSM-5-TR-guided filtering, Word2Vec skip-gram embeddings, Principal Component Analysis (PCA), L2 normalization, and K-Means clustering. A total of 42,212 raw posts were collected from five Indonesian Facebook mental health communities and filtered into 6,072 symptom-relevant documents. The optimal cluster structure was evaluated using the Elbow Method, Silhouette Score, Calinski-Harabasz Index, and Davies-Bouldin Index. The final K-Means model with K = 3 achieved a Silhouette Score of 0.5536, a Calinski-Harabasz Index of 11,360.55, and a Davies-Bouldin Index of 0.8472. The resulting clusters were interpreted as hopelessness and mental health crisis, emotional distress and social conflict, and anxiety and physical symptoms. Expert validation by a licensed psychologist supported the psychological relevance of these clusters. The findings indicate that Word2Vec, PCA, and K-Means can support symptomoriented exploration of Indonesian Facebook mental health narratives, while the resulting labels should be interpreted as thematic indicators rather than clinical diagnoses.