An evolutionary approach for depression detection from Twitter big data using a novel deep learning model with attention based feature learning mechanism

Abstract

One of the main factors causing suicide is depression. However, many cases of depression go undiagnosed because they are not correctly diagnosed. An increasing number of people with mental illnesses express their emotions online using tools like social media (SM) and specialized websites. Recently, efforts have been made to use Machine Learning (ML) and deep learning (DL) models to predict depression from SM platforms. However, it is problematic that most ML algorithms now provide no explanation. As a result, this study proposes a novel Deep Learning (DL) model called residual network 50, which includes optimal long short-term memory (RNT-OLSTM) for Depression Detection (DD) on Twitter data. In addition, to address the issue of data imbalance in the Twitter data, a cluster-based oversampling approach is used, which considerably reduces the possibility of bias towards the dominant class (non-depressed).. Finally, the embedding layers are inputted to RNT-OLSTM for DD, in which the hyperparameters of the network are tuned using the Sine Chaotic map and constriction factor-based Coyote Optimization Algorithm (SCCOA) to minimize the prediction loss. The out-comes prove that the proposed system performs better than the existing schemes for the DD of imbalanced Twitter data with higher detection rates.