Mental State Detection From Tweets By Machine Learning

Abstract

The world over, mental illness is a serious issue. Many people use the social media that may affect their mental health positively, but often result in negative sentiments. This research aims to determine an individual's mental state based on their social media behavior on Twitter. We analysed a dataset including 170000 real tweets by using natural language processing and machine learning techniques. Decision tree, support vector machine, and recurrent neural network (RNN) were used for classifying twitter users, to detect if they are in positive or negative mental state. These models were compared to determine which approach provides more accurate detection of a positive/negative mental state. Then, the RNN yielded the highest accuracy 0.76 among the models, with the precision, recall, and the F_1 score being 0.75, 0.74, and 0.75, respectively. The truncated singular value decomposition was also utilised to visualise the high-dimensional feature space of the data.