Stress Detection from Social Media Data by Graph Attention Network

The changes in human body in both mentally or physical will be considered as stress. Stress can be positive or negative. Positive stress helps us to motivate, alert during critical situations. For example during exam, a positive stress can be help to stay awake longer and might help our body to work harder. But if person go continues stress without relaxation and relief then it lead to health issue such as hair loss, heart attack and physiological problem. Stress can be diagnosed through counselling or medical practice. But in restricted lab environment it is difficult to express and measure correctly. The process is complex and lab experiment is costly. Stress and physiological related clinical process required more privacy. In today world people share their thoughts, feeling, emotion, current life status through social media. Social media data reflects the particular human life status in real-time. There are lot of Deep Learning and Natural Language processing available for text processing and sentiment analysis. Most of the modelsare rules based on the positive and negative score assigned to the words or phrase. The association between the words and grammatical structure is not considered for the prediction. The main goal of this project is to identify stress detection based on the Graph Attention Networks and Relational Graph Attention Networks. The sentence is constructed as Graph with wordsas nodes and relation between the words as edges. Social media datasets(SAD and Dreaddit) are preprocessed, trained and tested by GAT and RGAT. The results are compared with SVM and Multinomial Naive Bayesian. The accuracy for SAD and Dreaddit is 95 % and 77 % respectively. The accuracy is relatively higher than the expected accuracy provided in the data set publication.