A Novel adaptive learning vector quantization for time series classification

Abstract

Time series classification is a problem of interest in several areas of research, containing interesting applications for the use of machine learning techniques. Among the solutions adopted in the literature, the algorithms based on Artificial Neural Network (ANN) have been outstanding due to their generalization capacity. In this dissertation, a study was conducted on the performance of neural networks in the problem of time series classification. A new adaptive variation of the Learning Vector Quantization (LVQ) neural network, combined with a clustering method known as Self-Organizing Map (SOM), has been proposed. The proposed classifier, called Adaptive-LVQ-SOM (ALVQ-SOM), allows the removal and inclusion of prototypes in order to optimize the classification performance of the network. Two other methods inspiredby ALVQ-SOM are also presented: Driven-LVQ (dLVQ) and Driven-ALVQ-SOM (dALVQ).To evaluate the efficacy of the proposed method, a comparative study was conducted betweenthe classical LVQ classifiers, ALVQ-SOM and two other ANN-based classifiers: Multi-LayerPerceptron (MLP) and Support Vector Machine (SVM). In addition, the algorithmK- NearestNeighbors (k-NN) was inserted in this study, since this algorithm is considered a referenceclassifier in the literature of time series classification. The methodology adopted in the evaluationof the algorithms consists in the application of the cross-validation technique 10-fold in theexecution of simulations using different classifiers, applied to distinct datasets. The results ofthe experiments show that the proposed adaptive LVQ (ALVQ-SOM) method outperforms the classical versions of LVQ, presenting superior classification performance in most of the studieds cenarios