Efficient data abstraction using weighted IB2 prototypes

Data reduction techniques improve the efficiency of k-Nearest Neighbour classification on large datasets since they accelerate the classification process and reduce storage requirements for the training data. IB2 is an effective prototype selection data reduction technique. It selects some items from the initial training dataset and uses them as representatives (prototypes). Contrary to many other techniques, IB2 is a very fast, one-pass method that builds its reduced (condensing) set in an incremental manner. New training data can update the condensing set without the need of the “old” removed items. This paper proposes a variation of IB2, that generates new prototypes instead of selecting them. The variation is called AIB2 and attempts to improve the efficiency of IB2 by positioning the prototypes in the center of the data areas they represent. The empirical experimental study conducted in the present work as well as the Wilcoxon signed ranks test show that AIB2 performs better than IB2.