WINDOW SIZE THRESHOLD ANALYSIS FOR BRAINPRINT IDENTIFICATION USING INCREMENTAL K-NEAREST NEIGHBOUR (KNN)

• Main Authors: Siaw Hong, Liew, Yun Huoy, Choo, Yin Fen, Low, Shin Horng, Chong
• Format: Article
• Language: English
• Published: Asian Research Publishing Network (ARPN) 2020
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/39177/1/WINDOW%20SIZE%20-%20Copy.pdf; http://ir.unimas.my/id/eprint/39177/; http://www.arpnjournals.org/jeas/research_papers/rp_2020/jeas_0920_8305.pdf
• Institution: Universiti Malaysia Sarawak

This paper aims to investigate the window size threshold of the incremental update strategy in K-Nearest Neighbours (KNN) for brainprint identification. Electroencephalogram (EEG) signals are low signal-to-noise ratio and non-stationary. Incremental learning is good in handling dynamic applications. It does not require complete training examples; instead it is able to adapt dynamic changes to gradually form the target concept. KNN implements First-In-FirstOut (FIFO) strategy to guide the incremental learning updates. The FIFO strategy tends to construct the target concept from the training objects according to availability orders. If the number of training objects exceeds the predefined window size threshold, then the FIFO strategy remove the earliest available object. The step size of training pool is linear increased by 10%, from 20% up to 90%. The classification results showed improvement when the window size threshold is increasing. The optimum results recorded at the window size threshold of 60%, with 0.875 in accuracy, 0.887 in precision and 0.878 in f-measure. The degradation of the classification performance after 60% showed the FIFO incremental update strategy is less promising. Thus, future work should focus on the incremental update strategy for selecting the representative and distinct objects to improve the performance of brainprint identification.