Prototype generation based on MOPSO

Weiwei Hu,Ying Tan

DOI: https://doi.org/10.1049/pbce119h_ch1

2018-01-01

Abstract:When classifying a test instance, the nearest neighbor classifier will consume lots of time because it needs to search the whole training set for the instance's nearest neighbors. Prototype generation is a widely used approach to improve its time efficiency which generates a small set of prototypes to classify a test instance instead of using the whole training set. This paper applies the particle swarm optimization (PSO) to prototype generation and presents two novel methods to improve the classifier's performance. A fitness function named error rank is proposed to enhance the nearest neighbor classifier's generalization ability. In order to keep the classifier from overfitting the training set, this paper proposes the multiobjective optimization strategy which divides the whole training set into several subsets and regards the performance criterion on each subset as an objective function of the multiobjective PSO. The multiobjective optimization strategy pursues the performance over multiple subsets simultaneously, resulting in better generalization ability. Experimental results over 31 UCI datasets and 59 additional datasets show that the proposed algorithm achieves state-of-the- art performance.