Higher Order Neural Network and Metaheuristic Optimization based Machine Learning Approaches to Classification in Data Mining

Abstract

Modern system identification models are complex in such a way that proper handling of newlinenon-linear data is a complicated task. Various traditional system identification methods are newlineprojected by several researchers (Akaike s criterion; parametric prediction error methods), newlineand later on, a number of limitations related to these conventional approaches have been newlinenoticed (kernel-based approach and Gaussian regression approach for linear system newlineidentification). As a result, researchers are attracted towards system identification models newlinebased on hybridization of soft computing and machine learning techniques and found to be newlinesuccessful in determining optimal models. System identification models are widely used to newlineidentify models from the data in the field of science and engineering such as newlinebioinformatics, stocks, webs, e-commerce, transactions, scientific simulations, business newlinemanagement, microarrays gene expression, remote sensors, engineering designs and newlineproduction controls etc. In data mining, classification task can be visualized as a system newlineidentification problem, whose objective is to identify optimal model from the past data to newlineassign class labels to unknown patterns. Classification task determines unambiguous class newlinelabels by constructing a system identification model based on the past data, which is newlinehelpful to decide class labels of unknown patterns. It is often difficult to identify the newlineoptimal learning model for data classification problem as these real world data from newlinevarious domains are nonlinear in nature. Although, many conventional classification newlinemodels are proposed by several researchers based on statistics such as Bayesian classifier, newlineNearest neighbour classifier, C 4.5 classifier and Fuzzy decision tree etc., for the first time, newlineZhang et.al. (2000) have realized that artificial neural network (ANN) models may be the newlinesubstitute to those traditional classification models. ANNs are capable to identify intricate newlinemapping between the input and output data and therefore, the ANN based mode