Research Of Fine-Tuned Learning Clustering Algorithm

Clustering is an important technique which is used in various fields such as statistics for tremendous data,analysis of network,automatic supervised of medical images.Clustering divides data objects into different groups(clusters) according to the internal characteristics so that the elements attached to the different clusters have dissimilarities and the ones attached to the same clusters have similarities.In recent ten years,domestic or oversea researchers presented many clustering algorithms that mostly considered clustering in local or whole angle and tried to discover the excellent results from all clustering strategies.Owing to the large scale data and the complicated types,the clustering algorithms can't satisfy the people in the aspect of clustering quality.Meanwhile,classic algorithms are characterized by keeping the questions and the locality of data objects invariable.Aiming to improve the quality of clustering,this thesis adopts the characteristics of classic clustering algorithms,endeavors to analyses it in movable angle.Meanwhile,a new clustering algorithm based on Fine-Tuned Learning, which is called CAT-L,is proposed.The essence of CAT-L is to call a Fine-Tuned operator to construct a series of approximate problem of the original and the granularity of each approximate problem is different and depends on Fine-tuned factors.Firstly,some thin granularity approximate problems are clustered;secondly, thick approximate problems are constructed and solved by means of the results at first step.Since transformation changes the topology structure,the original problem can be converted to many simple problems and get to the destination that difficulty is reduced and clustering quality is improved.Classic clustering algorithm FCM and CLARANS are improved under the CAT-L framework.New clustering algorithms,FCM and CLARANS are executed on several synthetic dataset and three real dataset.Experiment results show that algorithm after betterment causes remarkable improvement on clustering quality under the condition of the constant time and space complexity.The new clustering algorithm presented in the thesis takes classic clustering algorithms as its attached ones and make them improved.This provides people with new ways and thoughts in the interest of acquiring better results and mining the hidden models and information.Thus,researches of this thesis have the great significance in theory and practice.