Detecting Signs of Depression Using Social Media Texts Through an Ensemble of Ensemble Classifiers

Artificial intelligence-based machine learning models have been widely used to explore and address various mental health-related problems in recent years, including depression. In this study, we present an ensemble approach to complement the 90 unique input features that we proposed in a previous study on depression detection using social media texts. Our proposed Ensemble of Ensemble Classifiers (EECs) combines many ensemble models, including Bagging Predictors, Random Forests, Adaptive Boosting and Gradient Boosting, as inner ensembles. These inner ensembles are arranged in a parallel fashion, where each of them is trained using different subsets of data sampled from the training data via bootstrap sampling. After the models are trained, during the testing phase, the results of all inner ensembles are processed using two methods—majority vote or class priority threshold—to get the final result as an output. From the experiments, we find that EECs are accurate in detecting signs of depression in social media users by analysing their posts in social media platforms such as Twitter. Our approach outperforms other ensemble methods on the public datasets we used. Moreover, if set correctly, the parameters of EECs can further improve the performance of the proposed ensemble in detecting signs of depression.