Clustering Algorithm Based On Mechanics

Clustering is a branch of data mining applied in many fields such as statistics of large data, analysis of network, automatic supervisory of medical images. Clustering is defined that data objects are divided to some groups according to the internal characteristics and satisfies the principle that maximal similarity in the same group and maximal dissimilarity between groups. In recent years, the researchers have proposed many clustering algorithms that observe clustering in the local or whole angle and endeavor to find all the ways to optimal results. Due to the data dealt are large-scale, complicated types, the present algorithms can't satisfy the applications. Meanwhile classic algorithms take only distance or density as the norm applied in clustering, so it is unreasonable and undesirable.For the sake of improving the quality of clustering, this thesis combines the advantages of classic algorithms, introduces mechanics and energy into clustering and present clustering algorithm based on mechanics. The principle of this algorithm is that the objects are taken as particles among which there exists gravity, particles are connected with elastic poles between which the hinge is used and the plane trusses are constructed. Trusses transform owing to outer power, when they get to steady state, the particles can be divided into clusters according to their displacement and the change of potential energy of the structures. This algorithm describes the relationship as power and optimizes the result according to energy. The advantages as followed: firstly, for power has the characteristics of combination and decomposition, the relationship among particles can be converted to the single ones equally, so it eliminates the deficiency that can't take the relationship among more than two particles into account at one time. Secondly, the equilibrium of particles will be broken thanks to power until it is set up again, the movement of particles shows the activity of gathering and scattering. Thirdly, energy is used to weigh the results and every cluster is considered as a unit which contains energy. The smaller is inner energy, the steadier is the structure. Experiments show that this algorithm has remarkable improvement on clustering quality.This thesis presents the clustering algorithm based on power and energy and provides the new ways and thoughts which guarantee that the accurate hidden information can be mined. So it is valuable in theoretical and practical fields.