Using Growing Neural Gas in Prototype Generation for Nearest Neighbor Classifiers

Instance-based learning algorithms, such as nearest neighbor NN classifiers, require storing all training instances and consulting them when making predictions. One alternative to overcome these costs is to reduce the learning dataset by a pre-processing step. This work deals with prototype generation, where new data points are generated from the original dataset. Reduction can be achieved by retaining less instances in the most representative areas of the dataset, which are represented by prototypes. Here Growing Neural Gas Networks are employed for generating the prototype instances. Experimentally, NN classifiers using the reduced datasets were able to maintain close accuracy to that of NN classifiers using the whole dataset.