Clustering Algorithm Based on Artificial Bee Colony Optimization

Abstract

After analyzing the disadvantages of sensitivity to the initial selection of the center, low clustering accuracy and the poor global search ability of k-medoids clustering algorithm, a clustering algorithm based on improved artificial bee colony (ABC) is proposed. By improving the initialization of bee colony, adjusting the search step dynamically with iteration increasing , and then introducing the selection probability based on sorting instead of depending on fitness directly, the ABC algorithm will quickly converge to global optimal. This paper will further optimize k-medoids to improve the performance of the clustering algorithm. The experimental results show that this algorithm can reduce the sensitive degree of the initial center selection and the noise, has high accuracy and strong stability. Introduction K-medoids algorithm is a classical algorithm to solve clustering problems [1]. Since it has the advantages in the convergence speed and local search ability, are widely used in data mining. However, it still have the problems of sensitive to initial centers, low clustering accuracy and poor global search ability. ABC [2, 3] algorithm is a novel swarm intelligence optimization algorithm by simulating the collecting nectar process of bee swarm which has few parameters, strong global search ability and robustness. An improved algorithm based on initialization center fine-tuning and incremental center candidate set was proposed [4]. The computational time get slightly lower, but clustering accuracy is not high. Combining granular computing to select the highest density K granules center as the initial clustering centers, the algorithm (GCK) has been a certain degree of optimization, but the initial clustering centers may be located in the same cluster [5]. Gao and Liu introduced fine-tuning mechanism into ABC algorithm and discussed the perturbation factor range, improved the local search ability [6]. Wang introduced forgetting factor and neighborhood factor into onlooker bees’ neighborhood search phase, enhanced the algorithm convergence [7]. In view of the above references is insufficient, an improved ABC algorithm which improves the colony initialization, dynamically adjusts neighborhood search step and introduces the probability based on sorting of onlooker bees is proposed. The improved ABC algorithm accelerate convergence speed and avoid premature convergence. Then we apply the improved ABC algorithm to optimize k-medoids. The experimental results show the proposed algorithm has faster convergence speed, higher accuracy and more stable performance. Granular Computing From the perspective of information granularity, the clustering analyze and solve problems under a uniform granularity [8]. Definition 1 Given the universeU , the partition of the knowledge P in U is{ } 1 2 , , , n X X X  , the particle density is defined in Eq. 1. ( ) gd X X U i i = . (1) International Conference on Applied Science and Engineering Innovation (ASEI 2015) © 2015. The authors Published by Atlantis Press 126 Where i X is the number of objects in the ith divided block, and U is the total object number of U . Suppose the number of particles is n , the average density of the particle is given in Eq. 2.