Improved sparse prototyping for relational K-means
Safouane Cherki, Parisa Rastin, Guénaël Cabanès, Basarab Mateï · 2016
This study investigates the possibility of improving K-means algorithm for non-vector data. It calls attention to the changes a similarity measure enforce compared to vectorial K-means, and aims to understand and take advantage of the opportunities offered by sparse prototyping for K-means. We propose here a new algorithm of clustering for relational data, i.e. data described by their relations to each other (usually their similarities). This algorithm computes a set of sparse prototypes to represent the data structure. The results are promising: the clustering quality of the sparse variation is similar to the traditional K-means, and the processing cost for high dimensions is lower than other relational algorithms.