AutoNLP for Optimal Number of Epochs in Multi-labeled Deep-Learning Models for Predicting Mental Disorders

This paper is part of our research doctoral thesis in which we are interested in developing efficient deep-learning models for predicting mental disorders using Twitter data. In particular, the aim here is to adopt hyperparameter optimization methods for the automatic selection of the optimal epoch number. Deep learning approaches are widely applied to various classification tasks. In particular, Recurrent Neural Networks (RNNs) have been shown to be effective in solving several problems related to natural language processing dealing with large volumes of unstructured textual data. Even so, these approaches are resource and time consuming due to their complexity and huge dimensionality. In this work, we are interested in multi-labeled deep-learning models for predicting depression and anxiety disorders using multi-labeled data extracted from the Twitter platform (known currently as the X platform), with a particular focus on the CNN-BiGRU hybrid model due to its performances over other models [14]. The aim is to involve hyperparameter optimization methods (HPO) to automatically tune the considered deep learning model and streamline the training process. The idea consists in using both the grid search hyperparameter optimizer and the early stopping strategy as conditions for a minimum loss, in order to retain the optimal epoch number for which the model is the well-performing throughout the whole training process. To validate our work, we consider two datasets namely Eval_Shen and Eval_Kaggle in addition to our own dataset. Overall, the experimental results show that the HPO based on early stopping and grid search could achieve some improvements by retaining the well-performing model of the optimal epoch number.