Using classifier fusion to improve the performance of multiclass classification problems

摘要

The problem of multiclass classification is often modeled by breaking it down into a collection of binary classifiers, as opposed to jointly modeling all classes with a single primary classifier. Various methods can be found in the literature for decomposing the multiclass problem into a collection of binary classifiers. Typical algorithms that are studied here include each versus all remaining (EVAR), each versus all individually (EVAI), and output correction coding (OCC). With each of these methods a classifier fusion based decision rule is formulated utilizing the various binary classifiers to determine the correct classification of an unknown data point. For example, with EVAR the binary classifier with maximum output is chosen. For EVAI, the correct class is chosen using a majority voting rule, and with OCC a comparison algorithm based minimum Hamming distance metric is used. In this paper, it is demonstrated how these various methods perform utilizing the Bayesian Reduction Algorithm (BDRA) as the primary classifier. BDRA is a discrete data classification method that quantizes and reduces the dimensionality of feature data for best classification performance. In this case, BDRA is used to not only train the appropriate binary classifier pairs, but it is also used to train on the discrete classifier outputs to formulate the correct classification decision of unknown data points. In this way, it is demonstrated how to predict which binary classification based algorithm method (i.e., EVAR, EVAI, or OCC) performs best with BDRA. Experimental results are shown with real data sets taken from the Knowledge Extraction based on Evolutionary Learning (KEEL) and University of California at Irvine (UCI) Repositories of classifier Databases. In general, and for the data sets considered, it is shown that the best classification method, based on performance with unlabeled test observations, can be predicted form performance on labeled training data. Specifically, the best method is shown to have the least overall probability of error, and the binary classifiers have the least overall average quantization complexity.