Combining diversity measures for ensemble pruning

• Presents an ensemble pruning method (DivP) that combines diversity measures.
• The combination of diversity measures is advantageous to prune pool of classifiers.
• DivP uses graph algorithms as a tool to group similar classifiers.
• DivP obtains better results than literature methods, such as: AGOB, DREP and GASEN.
• DivP generates smaller final ensemble than state-of-the-art methods.

Multiple Classifier Systems (MCSs) have been widely used in the area of pattern recognition due to the difficult task that is to find a single classifier that has a good performance on a great variety of problems. Studies have shown that MCSs generate a large quantity of classifiers and that those classifiers have redundancy between each other. Various methods proposed to decrease the number of classifiers without worsening the performance of the ensemble succeeded when using diversity to drive the pruning process. In this work we propose a pruning method that combines different pairwise diversity matrices through a genetic algorithm. The combined diversity matrix is then used to group similar classifiers, i.e., those with low diversity, that should not belong to the same ensemble. In order to generate candidate ensembles, we transform the combined diversity matrix into one or more graphs and then apply a graph coloring method. The proposed method was assessed on 21 datasets from the UCI Machine Learning Repository and its results were compared with five state-of-the-art techniques in ensemble pruning. Results have shown that the proposed pruning method obtains smaller ensembles than the state-of-the-art techniques while improving the recognition rates.