Font Size: a A A

Research Of Key Techniques In Fuzzy Clustering Based On Objective Function

Posted on:2013-06-24Degree:DoctorType:Dissertation
Country:ChinaCandidate:D H ChenFull Text:PDF
GTID:1228330395457126Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Clustering analysis is used to handle classification problem by mathematicalmethods, and is an important part of non-supervised pattern classification in patternrecognition. In the recent20years, it has been developed drastically. Fuzzy clusteringalgorithm has become the focus of research these years because it can describe andanalyze the uncertain relation accurately. Among the various fuzzy clustering algorithms,fuzzy clustering based objective function has been widely applied in image processing,pattern recognition and computer vision, which is the most efficient and popularmethods.Aiming at the faultiness, even serious drawbacks of the available clusteringalgorithm for the application in data mining, by combining the modified particle swarmoptimization and support vector clustering method,the traditional cluster analysisalgorithms are systematically improved and innovated in this thesis.The emphasis is puton the definition and optimization method of the objective function for fuzzy clustering,and a new fuzzy clustering algorithm with support vector is proposed to handlehigh-dimensional data sets with arbitrary shapes,which extends the application rangesof cluster analysis.In addition,a novel fuzzy clustering validity function is developedfor data mining. The experimental results illustrate the effectiveness and goodperformance of the proposed new ideas and new methods on fuzzy cluster analysis, andwell applied in national technology support projects.In sum, the main research results achieved in this paper are given as follows:1. The objective function of fuzzy clustering algorithm has been improved. Apossibilistic C-means clustering algorithm improved has been proposed. This method isfirstly used to calculate the membership matrix of data pattern and clustering center byimproving the objective function of PCM algorithm so as to complete the particleencoding. Thus, the sensitivity of algorithm to clustering center is reduced, the issue ofclustering consistency can be avoided as well, and the accuracy of clustering can beimproved. In consideration of the fuzzy clustering algorithm that is based on objectivefunction is a local search algorithm, the introduction of particle swarm optimization canlargely improve the overall optimization capacity that has a very satisfying searchcapability and clustering effect.2. The support vector clustering is researched. SVM (support vector machine)structures the decision hypersphere based on the structural risk minimization principle,which makes the maximum class margin between data. A new method of support vector fuzzy clustering is proposed on this basis. This method has a better performancecompared to the traditional one. After solving the issue of quadratic programming, anoverall optimal solution can be guaranteed. Besides, it can deal with the data set that isin any shape and divide the clustering shapes that have overlapping regions. Besides,the high-dimensional data can be solved as well. The experimental results proved thefeasibility and effectiveness of this method.3. The initialized clustering center of fuzzy clustering is researched. A method thatis kin to initialization based on density function is proposed in allusion to the issue thatmost of the fuzzy clustering algorithms based on objective function should be given theinitial clustering center before hand. This method confirms the clustering center throughdensity function of sample distribution. And the simulation experiment indicates thatthis method can not only get a very good clustering effect to the large data set ofhigh-dimension, its calculated amount can also be effectively controlled.4. The effectiveness index of clustering is researched. The confirmation of theoptimum number of fuzzy clustering and the selection of fuzzy weighted parameter tofuzzy clustering algorithm are within the range of effectiveness research. Therefore, aneffectiveness algorithm based on the partition coefficient and similarity measurement isproposed. This method not only takes the distribution characteristics of data sets intoaccount, but also integrates the fuzzy partition coefficient so that it makes the result ofclustering very distinct. And it can used to the evaluation of the effectiveness of datasets. An evaluation function on fuzzy decision is designed to optimize the selection ofweighting exponent. And the experimental result testified their effectiveness.
Keywords/Search Tags:Fuzzy Clustering, Fuzzy c-Means, Possibilistic c-Means, ClusteringCenter, Fuzzy Clustering Validity, Support Vector, Particle SwarmOptimization
PDF Full Text Request
Related items