Posterior probability based ensemble strategy using optimizing decision directed acyclic graph for multi-class classification

Ensemble strategy is important to develop a decomposition and ensemble method for multi-class classification problems. Most existing ensemble strategies use predetermined and heuristic decision rules. In this work, we build up the decision rules by optimizing decision directed acyclic graph (ODDAG) with classical and fuzzy decision trees to ensemble the posterior probabilities of binary classifiers from one-vs-one (OVO) or one-vs-all (OVA) decomposition strategies for multi-class classification problems. Four widely used extensible algorithms and ten decomposition and ensemble methods incorporating four binary classifiers (BCs) have been tested on 25 data sets. The empirical results show that the methods based on ensemble strategies using ODDAG are among the top methods that achieve the best performance in terms of two different measures.