Kernel Penalized K-means: A feature selection method based on Kernel K-means

We present an unsupervised method that selects the most relevant features using an embedded strategy while maintaining the cluster structure found with the initial feature set. It is based on the idea of simultaneously minimizing the violation of the initial cluster structure and penalizing the use of features via scaling factors. As the base method we use Kernel K-means which works similarly to K-means, one of the most popular clustering algorithms, but it provides more flexibility due to the use of kernel functions for distance calculation, thus allowing the detection of more complex cluster structures. We present an algorithm to solve the respective minimization problem iteratively, and perform experiments with several data sets demonstrating the superior performance of the proposed method compared to alternative approaches.