A Ranking Distance Based Diversity Measure for Multiple Classifier Systems

Multiple classifier fusion belongs to the decision-level information fusion, which has been widely used in many pattern classification applications, especially when the single classifier is not competent. However, multiple classifier fusion can not assure the improvement of the classification accuracy. The diversity among those classifiers in the multiple classifier system (MCS) is crucial for improving the fused classification accuracy. Various diversity measures for MCS have been proposed, which are mainly based on the average sample-wise classification consistency between different member classifiers. In this paper, we propose to define the diversity between member classifiers from a different standpoint. If different member classifiers in an MCS are good at classifying different classes, i.e., there exist expert-classifiers for each concerned class, the improvement of the accuracy of classifier fusion can be expected. Each classifier has a ranking of classes in term of the classification accuracies, based on which, a new diversity measure is implemented using the ranking distance. A larger average ranking distance represents a higher diversity. The new proposed diversity measure is used together with each single classifier's performance on training samples to design and optimize the MCS. Experiments, simulations, and related analyses are provided to illustrate and validate our new proposed diversity measure.