Fast normalized cut algorithm based on self-organizing map

Recently, researchers are paying more and more attention on the study of the normalized cut algorithm which has a lot of useful applications in different kinds of areas, such as medical image, image process, data mining, pattern recognition, and so on. Although the normalized cut algorithm is very effective to handle different kinds of challenging datasets, its computational cost is very high, especially when the sizes of the datasets are large. In order to solve this limitation, we propose a fast normalized cut algorithm based on self-organizing map (FNCUT(SOM)) to perform clustering on large datasets. FNCUT(SOM) pays more attention to the representative feature vectors which are the weight vectors of the neurons in SOM, instead of considering all the feature vectors. Specifically, FNCUT(SOM) adopts the self-organizing map to perform fast clustering on the dataset at first. Then, the weight vectors of the neurons in SOM serve as a new dataset, and is used to construct a representative matrix. In the following, the normalized cut algorithm is adopted to partition the representative matrix and obtains the structure of the dataset. Finally, two assignment criteria are designed to distribute the feature vectors in the original dataset into the corresponding clusters. The experimental results show that FNCUT(SOM) is effective and efficient when applied to perform clustering on the real datasets in UCI machine learning repository.