We want to develop a machine learning model which will predict if a person is depressed from their use of tweets on social media. Sentimental analysis can be very useful to find out if a person has depression (based on their tweets) and cure it before someone gets into serious trouble.
Background Social media data provides valuable clues about physical and mental health conditions. Although the users may not aware with their health, analysing such valuable indicators help to predict the mental health status of the person. Studies have revealed that in addition to physical test, predictive screening methods can be used for number of mental health conditions like depression, addictions, post-traumatic stress disorder (PTSD), and suicidal tendency. Although considerable works are going on with this field, predictive health screeing with social media is still in starting stage. In this project, we tried to build state-of-art for predicting whether a person may have depression or not based on his/her tweets on the Twitter.
Data Collection and Observation The tweets dataset stored on Kaggle was taken as dataset for study.
Summary Statistics Different information present on the dataset like number of rows, columns, variables, etc. were first observed. Number of tweets were counted along with their result like depressed or not were calculated and visualized using simple pie chart.
Machine Learining Models We trained supervised machine learning models to differentiate between depressed and healthy sample based upon their tweets. At the begining the sentences were passed to BERT models and pooling layer to generate their embeddings and finally normalized. The Classifiers were trained on a randomly selected 80% of total population, and tested on the remaining 20%. We basically used Random Forest Classifier and Naives Bayes Classifier and Support Vector Machine (SVM). For each model the output was analysed as classification report, accuracy and infusion matrix. Out of several candidate algorithms, the SVM demonstrated the best performance (Accuracy of 99%). Moreover, Random Forest Classifier also showed good accuracy (95.9%). Finally the hyperparameter tuning was done with Random Forest Classificer and found the slightly better accuracy (96.2%). The whole codes which is running in the backend is connected to simple GUI where texting some tweets will show whether it is depression or not.