Identifying stable objects for accelerating the classification phase of k-means

Abstract

This work presents an improved version of the K-Means algorithm, this version consists in a simple heuristic where objects that remains in the same group, between the current and the previous iteration, are identified and excluded from calculi in the classification phase for subsequent iterations. In order to evaluate the improved version versus the standard, three synthetic and seven well-known real instances of specialized literature were used. Experimental results showed that the proposed heuristic spends less time than the standard algorithm. The best result was obtained when the Transactions instance was grouped into 200 clusters, achieving a time reduction of 90.1% regarding the standard version, with only a grouping quality reduction of 3.97%.