An Investigation of Ensemble-Based Framework for Stress Detection Using Multimodal Clues

Stress detection using multimodalities has emerged as a prominent area of research, attracting significant attention from researchers in recent times. The primary goal of this work is to classify individuals as either stressed or unstressed. For this, we experimented with different ensemble methods (bagging, boosting, and stacking) on a publicly available SWELL-KW dataset. The study addresses three research questions. Firstly, it investigates whether integrating multiple base models using ensemble methods improves the accuracy of stress predictions when compared with using single models. This evaluation aims to enhance the generalisability of stress prediction models. Secondly, the study explores the capability of ensembles to effectively integrate and fuse data from multiple modalities, leveraging the strengths of individual modality to improve overall predictive performance. By integrating multimodal data within an ensemble framework, the study aims to enhance stress prediction accuracy. Lastly, the impact of different ensemble approaches on stress prediction performance is examined. Through these investigations, the study contributes to the field by evaluating the effectiveness of ensemble methods, exploring the benefits of multimodal data integration, and determining optimal ensemble techniques for stress prediction.