Empirical comparison of four classifier fusion strategies for positive-versus-negative ensembles

Abstract

Classification modeling is commonly used for predictive data mining to create models (classifiers) that can predict the values of qualitative variables. Ensemble classification is concerned with the creation of many base classifiers which are then combined into one predictive classification model. Positive-versus-negative (pVn) classification has recently been proposed in the literature as an ensemble classification method with a potential to provide high predictive performance. Many methods of combining base model predictions for ensembles have been reported in the literature. The purpose of this paper is to report on a study that was conducted to compare four methods of combining base model predictions for pVn ensemble classification. The four methods that were studied are the max rule, min rule, sum rule and product rule. The four rules were studied for classification tree and artificial neural network pVn ensemble classification using a benchmark dataset for computer network intrusion detection systems. The main conclusion from the studies is that the sum, product and min rules provide predictive performance which is at least as high as that provided by the max rule for pVn classification.