ATISA: Adaptive Threshold-based Instance Selection Algorithm

• We propose three instance reduction techniques called ATISA1,2,3.
• ATISA maintain important border and inner points per class based on an adaptive threshold.
• When compared with the state-of-the-art algorithms, ATISA1 obtains better accuracy rates and promising reduction rates.
• ATISA is faster than DROP3, ICF and HMN-EI.

Instance reduction techniques can improve generalization, reduce storage requirements and execution time of instance-based learning algorithms. This paper presents an instance reduction algorithm called Adaptive Threshold-based Instance Selection Algorithm (ATISA). ATISA aims to preserve important instances based on a selection criterion that uses the distance of each instance to its nearest enemy as a threshold. This threshold defines the coverage area of each instance that is given by a hyper-sphere centered at it. The experimental results show the effectiveness, in terms of accuracy, reduction rate, and computational time, of the ATISA algorithm when compared with state-of-the-art reduction algorithms.