Research On Clustering Algorithm Based On Extreme Learning Machine

Clustering technology is a critical method in data analysis, and it occupies an even more important place with the explosion of information. People incorporate their experiences into the clustering technology by using their summarizing and concluding on data. Use computers to find inherent patterns from the data it selves is a significant conversion from human wisdom to artificial intelligence. Artificial neural networks is a great learning model that it can bring human knowledge to computers and a fusion of artificial neural networks and clustering methods will be improving clustering performance. Taking the above issues into consideration, the research content of this paper is as follows:Firstly, this paper reviews the development process of clustering technology and artificial neural networks. By summarizing the characteristics from different kinds of clustering methods. We detect the consistence of artificial neural networks and clustering technology. After that, a fast learning model, which is called Extreme Learning Machine, is picked out to have potential advantages in clustering by union with classical clustering method called K-means.Secondly, we summarize the framework of clustering algorithms based on Kmeans algorithm, and present a modular description of the clustering framework which concludes feature mapping, clustering center initialization, similarity functions, update method of clustering center and termination conditions. And we also provide several alternative implementation methods for each modular in the third chapter.Thirdly, a new clustering algorithm is proposed based on Extreme Learning Machine and K-means clustering algorithm, we named it K-Extreme Learning Machine Clustering, and is abbreviated to KELMC. For further improve the robustness of the proposed method, we upgrade the KELMC algorithm with the illumination of PCA and manifold, and generate another two clustering named as EP-KELMC and MF-KELMC.Finally, we evaluate our algorithms on six toy datasets and six UCI datasets. In order to find properties of the proposed methods, we also discussed the rules of the accuracy fluctuate varies to different parameters. From the experiment results, proposed methods outperform state-of-the-art clustering algorithms on most data sets.