A Research On Attribute-Weighting Cluster Algorithm

Human beings recognize the world according to make the distinctions among various of things and be familiar with the similarities with different things. Clustering analysis is a multivariate analysis method which belongs to mathematical statistics to study"things of one kind come together", clustering analysis is widely applied in various of fields such as data mining, pattern recognition, management of the economy and so on. Currently, k-means algorithm, transitive closure and Iterative self-organizing data analysis techniques are the typic clustering algorithms. The importance of each attribute is always regarded as the same in these algorithms, and the attributes are granded the same weights. In an actual cluster problem, the contribution of each attribute is always not the same, the important attributes have priorities. Therefore, target to the clustering algorithm based on attribute weight learning, the chief research tasks are as fellows:In order to reflect the contribution of each attribute is not the same in the clustering, this paper researches the existing two methods to gain the attribute weights, they are the entropy to get the attribute weights for the clustering algorithm and the gradient descent algorithm to get the attribute weights for the clustering algorithm; And based on the advantages of the gradient descent algorithm to get the attribute weights for the clustering algorithm, this paper puts forward the modified particle swarm optimization algorithm to learn the attribute weights. This paper researches the traditional clustering algorithm and attribute-weight clusteiring algorithm. Extensive experimental results on UCI machine learning repository indicate that the performance of the attribute-weight clustering algorithm is better than the traditional clustering, and attribute-weight clusteiring algorithm based on modified particle swarm optimization algorithm avoids the local optimum solutions, so this method can further improve the performance of the clustering algorithm with the more appropriate attribute weights.