A STACKED ENSEMBLE TECHNIQUE WITH GLOVE EMBEDDING MODEL FOR DEPRESSION DETECTION FROM TWEETS

Abstract

With the increasing volume of web content on social network sites like Facebook, Twitter, etc. identifying the attitude of people becomes an easy task. That attitude can be used as an input to find the mental status of that person through their texts. According to WHO, depression is a general mental disorder, which has already affected more than 264 million people. With the help of sentiment analysis, it is possible to detect depression at an early stage from their tweets as they represent their attitude. Machine Learning Classification algorithms help to classify the texts as Depressed or non-depressed, but their accuracy is limited when researchers are using only traditional Bag of Words vectorizers to extract features. Instead of this, word embedding models can be used which represent words as real-valued vectors in a distinct vector space that is already defined and provides better accuracy. In this paper, we try to implement eight machine learning techniques for depression detection from tweets namely SVM, Logistic Regression, ExtraTree, Bagging, Random Forest, AdaBoost, XG boosting, and Gradient Boosting which employs different word embeddings like Word2Vec, FastText, and Glove on the dataset which is used to train the classifiers and performed a comparative evaluation of word embedding models with different classifiers. This study proposes a new Stacked ensemble technique with a glove embedding model for Depression detection from Tweets which provides higher accuracy than standalone models.