Growing generalized learning vector quantization with local neighborhood adaptation rule

Prototype based learning algorithms, such as Kohonen´s learning vector quantization (LVQ) algorithm and its variants, offer the simple and intuitive model while excellent generalization performance in pattern classification tasks. As one of the powerful variants of the LVQ, the generalized LVQ (GLVQ) algorithm has shown promising performance in many applications. However, the convergence of the GLVQ algorithm heavily depends on the initializations. Furthermore, it is hard to reasonably assign the number of labeled prototypes to different classes in advance due to lack of knowledge about the characteristics of the training set. We present a novel growing generalized LVQ (G-GLVQ) algorithm. Through combining a local neighborhood adaptation rule devised by us for the GLVQ training with the growth procedure inherited from the growing neural gas, our proposed G-GLVQ algorithm can be insensitive to the initial prototypes position and avoid subjectively predefining the number of prototypes for each class before training. As training proceeds, a newly generated prototype can be assigned with the proper class label and inserted at a suitable position. In addition, the topological relations among all prototypes can be established automatically. Experimental results on artificial multimodal type and UCI datasets have demonstrated the superior classification accuracy and stability of our algorithm than the original GLVQ and one of its variants.