On ensemble learning for mental workload classification

Tools

McGuire, Niall and Moshfeghi, Yashar; Nicosia, Giuseppe and Ojha, Varun and La Malfa, Emanuele and La Malfa, Gabriele and Pardalos, Panos M. and Umeton, Renato, eds. (2024) On ensemble learning for mental workload classification. In: Machine Learning, Optimization, and Data Science. Lecture Notes in Computer Science . Springer-Verlag, GBR, pp. 358-372. ISBN 9783031539664 (https://doi.org/10.1007/978-3-031-53966-4_27)

[thumbnail of McGuire-Moshfeghi-Springer-2024-On-ensemble-learning-for-mental-workload-classification] Text. Filename: McGuire-Moshfeghi-Springer-2024-On-ensemble-learning-for-mental-workload-classification.pdf
Accepted Author Manuscript
Restricted to Repository staff only until 15 February 2025.
License: Strathprints license 1.0
Download (746kB) | Request a copy

Abstract

The ability to determine a subject’s Mental Work Load (MWL) has a wide range of significant applications within modern working environments. In recent years, techniques such as Electroencephalography (EEG) have come to the forefront of MWL monitoring by extracting signals from the brain that correlate strongly to the workload of a subject. To effectively classify the MWL of a subject via their EEG data, prior works have employed machine and deep learning models. These studies have primarily utilised single-learner models to perform MWL classification. However, given the significance of accurately detecting a subject’s MWL for use in practical applications, steps should be taken to assess how we can increase the accuracy of these systems so that they are robust enough for use in real-world scenarios. Therefore, in this study, we investigate if the use of state-of-the-art ensemble learning strategies can improve performance over individual models. As such, we apply Bagging and Stacking ensemble techniques to the STEW dataset to classify “low”, “medium”, and “high” workload levels using EEG data. We also explore how different model compositions impact performance by modifying the type and quantity of models within each ensemble. The results from this study highlight that ensemble networks are capable of improving upon the accuracy of all their individual learner counterparts whilst reducing the variance of predictions, with our highest scoring model being a stacking BLSTM consisting of 8 learners, which achieved a classification accuracy of 97%.

ORCID iDs

McGuire, Niall and Moshfeghi, Yashar ORCID logoORCID: https://orcid.org/0000-0003-4186-1088; Nicosia, Giuseppe, Ojha, Varun, La Malfa, Emanuele, La Malfa, Gabriele, Pardalos, Panos M. and Umeton, Renato
CORE (COnnecting REpositories)