A Hybrid Ensemble Method for Multiclass Classification and Outlier Detection

Multiclass problem has continued to be an active research area due to the challenges paused by the issue of imbalance datasets and lack of a unifying classification algorithms. Real world problems are of multiclass nature with skewed representations. The study focused on the challenges of multiclass classification. Multiclass datasets were adopted from UCI machine learning repository. The research developed a heterogeneous ensemble model for multiclass classification and outlier detection that combined several strategies and ensemble techniques. Preprocessing involved filtering global outliers and resampling datasets using synthetic minority oversampling technique (SMOTE) algorithm. Datasets binarization was done using OnevsOne decomposing technique. Heterogeneous ensemble model was constructed using adaboost, random subspace algorithms and random forest as the base classifier. The classifiers built were combined using average of probabilities voting rule and evaluated using 10 fold stratified cross validation. The model showed better performance in terms of outlier detection and classification prediction for multiclass problem. The model outperformed other commonly used classical algorithms. The study findings established proper preprocessing and decomposing multiclass results in an improved performance of minority outlier classes while safe guarding integrity of the majority classes.