Class Representatives Selection in non-metric spaces for nearest prototype classification

The nearest prototype classification is a less computationally intensive replacement for the $k$-NN method, especially when large datasets are considered. Centroids are often used as prototypes to represent whole classes in metric spaces. Selection of class prototypes in non-metric spaces is more challenging as the idea of computing centroids is not directly applicable. Instead, a set of representative objects can be used as the class prototype.

This paper presents the Class Representatives Selection (CRS) method, a novel memory and computationally efficient method that finds a small yet representative set of objects from each class to be used as a prototype. CRS leverages the similarity graph representation of each class created by the NN-Descent algorithm to pick a low number of representatives that ensure sufficient class coverage. Thanks to the graph-based approach, CRS can be applied to any space where at least a pairwise similarity can be defined. In the experimental evaluation, we demonstrate that our method outperforms the state-of-the-art techniques on multiple datasets from different domains.