Research On Hybrid Clustering Algorithm Based On Particle Swarm Optimization Algorithm

Clustering is an important data analysis technique.Research shows that no clustering algorithm can achieve the best clustering performance in all application scenarios.Therefore,it is a meaningful work to propose and improve different types of clustering algorithms to meet the needs of different scenarios.The partitioning methods are classic clustering methods.They are simple and efficient,and have been widely used in various fields.However,many partition clustering methods,such as K-means,K-medoids,and Fuzzy C-means,still have problems that need to be solved:(1)It is very sensitive to the selection of the initial cluster centroids,and improper initial cluster centroids may lead to unsatisfactory clustering results;(2)The objective function is often a non-convex function containing multiple local extremum,which will lead to the local optimum when minimizing the objective function;(3)Poor stability.Since the clustering task can be modeled as an optimization problem,it is a natural choice to use intelligent optimization algorithms to solve the clustering problem.The particle swarm optimization(PSO)algorithm is an effective global optimization algorithm.It guides the particles to search for the optimal solution in the complex data space by simulating the foraging process of birds in nature.In recent years,the PSO algorithm has been considered an effective method to solve the clustering problem.However,the basic PSO also has the possibility of converging to a local optimum,especially when solving complex problems.In order to solve the above problems,this paper first improves the basic PSO,and then combines the improved PSO with K-means to propose a hybrid clustering algorithm.The main research contents and contributions of this paper are as follows:(1)Propose the GLPSO algorithm based on Gaussian distribution estimation method(GEDM)and Lévy flight strategy.GEDM can use the promising particles in the population to better estimate the evolution direction of the population and thus speed up the convergence speed.The Lévy flight strategy can generate disturbances when the population is stuck in a stagnant state.In addition,the GLPSO algorithm also adopts boundary control strategies to limit the search range and flight speed of particles,and adopts a greedy strategy to retain promising particles from parent particles and newly generated candidate particles.The experimental results on the CEC 2014 benchmark test function set show that GLPSO has stronger optimization capability than the basic PSO and the four advanced PSO variants.(2)Combining GLPSO and K-means,a hybrid clustering algorithm GLPSOK is proposed by making full use of the simplicity and high efficiency of K-means and the powerful global search ability of GLPSO.In the GLPSOK algorithm,a small number of particles in the initial population are initialized with the clustering results of K-means to ensure that there are some relatively good particles in the initial population,and the remaining particles are initialized randomly to increase the initial population's diversity.The experimental results on six synthetic data sets and five UCI real data sets show that the hybrid clustering algorithm GLPSOK proposed in this paper has better performance and stronger stability than eight classical or advanced clustering algorithms.