Examining Variants of Learning Vector Quantizations According to Normalization and Initialization of Vector Positions

Abstract

Learning Vector Quantization is a prototype-based artificial neural network. The classification is performed by representing the data set with the prototype vectors of the classes. In this study, using some variants of Learning Vector Quantization such as LVQ1, LVQ2.1, LVQ3, LVQX, and OLVQ1, the systems are designed and implemented, and they are examined according to initializations of prototype vectors and data sets. Every data set is divided into training and testing data sets. With the training data set, all LVQ networks are trained in a reinforcement learning strategy, and the models for each network are generated to test the success of the systems. In addition, the systems are compared with each other using some distinct normalization techniques such as z-score and linear scaling. In initial conditions, all prototype vectors can be randomly selected, and the values of all prototype vectors can be assigned to zero. The generated systems are evaluated by accuracy and f-measure benchmark measures and compared by their success rates.