A comparative work of incremental learning and ensemble learning for brainprint identification
|
Full Text |
Pdf
|
|
Author |
Siaw-Hong Liew, Yun-Huoy Choo, Yin Fen Low and Fadilla Atyka Nor Rashid
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
1249-1257
|
|
Volume No. |
18
|
|
Issue No. |
11
|
|
Issue Date |
August 13, 2023
|
|
DOI |
https://doi.org/10.59018/0623161
|
|
Keywords |
incremental learning, ensemble learning, electroencephalogram (EEG) signals, brainprint identification.
|
Abstract
Electroencephalogram (EEG) signals are nonstationary and vary across time. The static learning model requires large training data to ensure sufficient knowledge acquisition to build a robust model. However, it is very challenging to achieve complete concept learning due to the behavioural changes in model learning. This issue is particularly critical in brainprint identification, where data acquisition in a short time cannot ensure sufficient training data for comprehensive model learning. Thus, dynamic learning, i.e., incremental learning and ensemble learning, presents a better solution for encapsulating EEG signal changes and variations. Both incremental and ensemble learning follow different approaches to manag the concept learning. Incremental learning merges new variations of EEG signals into the existing learning model over time, while ensemble learning uses multiple models for prediction. Nevertheless, limited research works were reported on comparing these two learning methods to prove the efficiency in handling nonstationary data for brainprint identification. Thus, this paper aims to compare incremental learning and ensemble learning for brainprint identification modelling. Incremental Fuzzy-Rough nearest Neighbour (IncFRNN) and Random Forest are selected to represent incremental learning and ensemble learning, respectively. Accuracy, area under the ROC curve (AUC) and F-measure were used to evaluate the classification performance. The experimental results proved that incremental learning outperformed ensemble learning when the training data were limited. The classification results of IncFRNN model were recorded at 0.9160, 0.9827 and 0.9169 while the Random Forest model only yielded 0.8113, 0.9709, and 0.9169 in accuracy, AUC, and F-measure, respectively. The ongoing learning process in incremental learning helps to capture the new changes in EEG signals and improve the classification performance.
Back