Dynamic Prototype Addition in Generalized Learning Vector Quantizers
The classification accuracy achieved by a statistical machine learning algorithm is determined by selecting the function(s) that most closely match a domain that results in the correct labeling of new samples, while not over-fitting. This is termed the accuracy/generalization trade-off. In Learning Vector Quantization (LVQ) the underlying function is a set of vectors in the domain space that have a relationship with each other and are prototypical of the underlying data. Where the prototypes are and how many there are influences the piece-wise linear decision boundaries they create and hence the final accuracy of the resulting LVQ. This work develops a novel framework that includes the LVQ learning components of Competition, Winner Selection, and Synaptic Adaptation, and adds a new component Network Structure Modification (NSM) that allows for experimentation of four different prototype addition strategies: simple, cost-minimizing, clustering, and cost-minimizing/clustering hybrids. Within these strategies, seven novel methods are tested on data sets using the Generalized Relevance LVQ Improved (GRLVQI) algorithm. Results show that cost-minimizing strategies achieve the highest classification accuracy but do not generalize as well, clustering methods tend to generalize well but are less accurate, and hybrid strategies provide the best trade-off between accuracy and generalization.