Selection And Integration Algorithm Of Eigenvector In Spectral Clustering

Spectral clustering algorithm has a good performance on sample space in arbitraryshapes. In recent years, spectral clustering has received lots of attention in datamining, machine learning, etc. Selective ensemble has become a popular researchproject due to reducing the storage requirements, improving the prediction speed andthe performance of ensemble learning machine. On the basis of spectral clustering,this paper adopted selective ensemble methods in order to achieve better clusteringresults.Existing studies showed that the first k eigenvectors may have no desiredperformance in spectral clustering, it was necessary to select eigenvectors fromscratch, and a given eigenvector group was not necessarily suited to the dataset, soselective ensemble for eigenvector groups was taken into consideration in this paperto improve the clustering performance, including the evaluation of eigenvectors, thegeneration of base eigenvector groups and selective ensemble strategy.Firstly, an eigenvector selection algorithm in spectral clustering based on Baggingwas proposed in order to select better eigenvectors. The Constraint Score was used toevaluate the eigenvectors of training data set in order to select better eigenvectors, andthen the selected eigenvectors under different constraints were integrated into bettergroups of eigenvectors by Bagging strategy. To test the obtained eigenvector groups,spectral clustering was carried out on the corresponding eigenvectors of testing dataset. Experimental results showed that the proposed algorithm could gain satisfactoryprediction results.Secondly, a dynamic selective ensemble algorithm of eigenvectors for spectralclustering was proposed. The aforementioned Bagging selection method was used,and some good base eigenvector groups were obtained to distinguish between the datain the same class and the data in different classes effectively. Generally, the baseeigenvector groups with strong discrimination ability appeared frequently. And thenfor each testing datum, the clustering accuracy of its l-nearest neighbors from trainingdataset was used to evaluate eigenvector groups, several accurate eigenvector groupswere selected to vote. To test the obtained eigenvector groups, spectral clustering wascarried out on the corresponding eigenvectors of testing data set. The clusteringresults were aligned and the final experimental results were obtained. Thediscrimination performance of eigenvector group is taken into account in the selectionof eigenvector groups, and the eigenvector groups used for voting were different fromeach other to keep the ensemble diversified. At the same time, the clusteringperformance of eigenvector groups on l-nearest neighbors of testing data was takeninto consideration to maintain the accuracy of the ensemble. Experimental resultsshowed that the proposed dynamic selective ensemble algorithm could improve theclustering performance of testing data.Finally, experimental methods were used to study the relationship betweenaccuracy and diversity in ensemble learning, and to determine whether there was acertain correlation between them. In conclusion, this paper mainly studied the selection methods and ensemble methods of eigenvectors in spectral clustering, andexperimental results showed that the designed selection methods and ensemblestrategies were effective.