Model Comparison for the Classification of Comments Containing Suicidal Traits from Reddit via NLP and Supervised Learning

Abstract

In recent years, suicide has become one of the most critical issues regarding public health between teenagers and adults. On the other hand, the growth and wide-spread of social networks and mobile devices have allowed us to compile relevant information that helps us understand the thoughts, feelings, and emotions extracted from these platforms. The detection of suicidal traits on social media has be-come one relevant research topic. It has permitted the identification of probable suicide traits among media users by examining their posts on known social net-works such as Reddit. For that reason, the purpose of the present research is to compare different supervised classification models such as Logistic Regression, Support Vector Machines, Random Forest, AdaBoost, Gradient Boosting, and XGBoost; together with feature extraction techniques such as TF-IDF and Glove. The results from our experiments show that the best model is SVM with TF-IDF obtaining metrics of 91.50% in Accuracy, 92.40% in Precision, 90.30% in Re-call, and 91.50% regarding the F1-score. This study also shows that TF-IDF for feature extraction outperforms Glove when applied to the different models tested.