Research On Swarm Intelligence Based Multiobjective Clustering Algorithms

Clustering is an unsupervised algorithm, which has been widely researched in many fields, such as data mining, pattern recognition and image processing. During the recent several decades, various kinds of clustering algorithms were proposed, among which the objective funciton based clustering algorithms were the most popular. Generally, these algorithms have many advantages such as ease to be designed and understood, wide application. However, these algorithms usually fall into local optimum, and mostly only optimize a single criteria, which suffer from many limitations. This paper studies clustering algorithms based on swarm intelligent and multiobjective optimization, and the main research contributions are listed as follows.(1) The K-harmonic means algorithm(KHM) has a disadvantage to easily fall into local optimum, hence a hybrid KHM based on improved firefly algorithm(FA) was proposed to solve it. The parallel chaotic based elaborate searching were utilized to search the nearby region of the current best and second best solutions found by FA, in order to enhance the searching ability of the algorithm. And then the improved FA was used to optimize the centers of clusters which were obtained by KHM. In addition, another disadvantage of KHM is to view all features as the same importance when calculating distance between the points and centers, hence an improved clustering algorithm is proposed which takes advantage of feature weighting during the clustering procedure. In the objective function of the algorithm, the Euclidian distance is replaced by the weighted Euclidian distance, and the update procedure of feature weight to ensure the convergence is proved. Experimental results show that chaotic FA can help KHM algorithm escape from being trapped into local optimum effectively, but the improvement degree is not so obvious. Moreover, the performance of these types of algorithms is improved largely with feature weighting.(2) The existing attribute weighted clustering algorithms almost optimize only a single objective function cannot effectively cope with different kinds of datasets, and the performance is not very desirable for complex datasets with non-hyper spherical shapes and/or linearly non-separable patterns. In this paper, the multiobjective optimization strategy is utilized to improve kernel-based attribute weighted clustering algorithm, where two objective functions that separately considered within-cluster and between-cluster information were optimized simultaneously. And a sampling operation and efficient clustering ensemble are incorporated to improve the projection coordinate based method of achieving the clustering solution, which can reduce the time complexity especially for relatively large datasets. The experiment results show that, the clustering accuracy and stability of the proposed algorithm is superior to the existing attribute weighted algorithms and the computing efficiency is improved by a large margin.(3) It has been shown that most traditional clustering algorithms for categorical data that only optimize a single criteria suffer from some limitations, thus a novel multiobjective fuzzy clustering is proposed, which simultaneously considered within-cluster and between-cluster information. The lately reported algorithms are all based on K-modes, and the more accurate algorithm fuzzy centroids is utilized as the base algorithm to design the proposed method. Fuzzy membership is used as chromosome that is different from traditional genetic based hybrid algorithms, and a set of optimal clustering solutions can be produced by optimizing two conflicting objectives simultaneously. Meanwhile, a termination criterion in advance which can reduce unnecessary computing cost is used to judge whether the algorithm is steady or not. To further improve the efficiency of the proposed method, fuzzy centroids can be calculated using a subset of the dataset, and then the membership matrix can be calculated by these centroids to obtain the final clustering result. The experimental results show that the clustering accuracy and stability of the proposed algorithm is better than the state of art multiobjective algorithm, and also the computing efficiency is improved to a large extern. Large superiority has been achived over similar algorithms.