Improving multiclass classification and outlier detection method through ensemble technique

Abstract

Class imbalance problems in multiclass have attracted much research focus due to classification difficulty caused by imbalance class distribution, presence of outliers, and irrelevant features that degrades performance of classifiers. Most of the commonly used classification algorithms are developed for binary problem. Rare classes in the multiclass datasets may be treated as outlier classes. The study proposed development of an ensemble multiclass classification and outlier method using adaboost, random subspace algorithms with random forest as base classifiers and voting combination technique. Experimental results shows that the proposed method outperformed most of the commonly used classification algorithms such as KNN, bagging, Naive Bayes, and random forest. Using rare classes as the outlier classes, the proposed method performed better than outlier detection methods built using Naive Bayes, KNN, decision trees, and random forest algorithms. The study concludes that ensemble techniques leads to an improved multiclass classification and outlier detection method.