Swarm Intelligence Based Clustering Analysis

Clustering analysis is one of the most fundamental research areas in data mining.Its objective is to divide the data sets into individual clusters based on the similarity.With data becoming more complex in big data era,new clustering methods are put forward under the calling for faster processing speed and clustering accuracy.Those based on Swarm Intelligence(SI)seems much more popular among these new clustering methods,for its robustness,spontaneity and potential parallelism.SI simulates the social behavior of swarms in the nature,like birds,fish,ants and so on.And the concept has been wildly implemented in many applications.The thesis mainly consists of three parts,shown in below.1)Background knowledge introduction.Starting from the outline of the development of clustering analysis,followed by the history of SI,clustering algorithms based on SI were mainly discussed including the basic principles,algorithms,pros and cons,and related modified algorithms.2)A Subtractive Particle Swarm Optimization(SCPSO)algorithm has been proposed.SCPSO adopted Subtractive Clustering(SC)algorithm to initialize the particles in PSO,which helps swarm improve the clustering efficiency and purity index.As experiments shown,SPSO could improve the performance of clustering in some extend.3)An Adaptive Quantum-behaved Particle Swarm Optimization with Crossover Operators(AQPSO CO)has been proposed,a novel algorithm for clustering data based on Quantum-behaved Particle Swarm Optimization.The algorithm mainly consists of two steps.First,a new real-time contraction-expansion coefficient update function is adopted to distinguish the two main phases,which are convergence phase and exploration phase during searching in solution space.At last,crossover operators are employed to retaining probability of better dimensions information,meanwhile avoiding the stagnation in movement of the particle.In order to testify the cluster performance of our approach,our approach is evaluated on 3 data sets,and compared to the performance of the K-Means+PSO,the original QPSO.Results show that our approach(ACQPSO)statistically outperforms the others in terms of convergence speed,and the evaluation of fitness.