Hybridization of Basic Classification Algorithms Based on Instance Filtering

Abstract

The quality of training data has a considerable influence on the learning results of a basic classification algorithm A model induced by a basic classification algorithm generally exhibits a high degree of instability and limitations Ensemble algorithms that produce a set of models by employing one or more basic classification algorithms are proposed to resolve the deficiencies of basic classification algorithms However when a prediction made by the majority vote of a set of models is difficult to interpret and the training cost of the models is relatively high A hybrid classification algorithm integrates basic ones for data preprocessing and class prediction Misclassified instances are generally considered as noise and thus excluded from learning However the excluded data may contain useful information in classifying some new instances This study proposes hybrid classification algorithms based on instance filtering and each one of them is a combination of two basic algorithms One plays the role of instance filtering and the other is to build three classification models Every new instance will be classified by only one of the three models and hence the interpretation of every prediction remains easy Na?ve Bayesian classifier and decision tree induction are the two basic algorithms for composing hybrid ones to process discrete data and the hybrid algorithms for continuous data are composed of k-nearest neighbor and support vector machine The hybrid classification algorithms are tested on 20 data sets to demonstrate that they can outperform basic algorithms and the hybrid algorithm proposed by a previous study

Date of Award	2020
Original language	English
Supervisor	Tzu-Tsung Wong (Supervisor)