Local within-class accuracies for weighting individual outputs in multiple classifier systems

A new method for assigning weights to individual classifiers in a multiple classifier system based on their local within-class accuracies is proposed. First distance metric learning is applied to determine the within-class nearest neighbors for an example to be classified. Then the local within-class accuracy for an individual classifier while classifying this example is judged by its performance on these neighbors. Experiments on a number of data sets with comparisons to two other existing methods show the effectiveness of the proposed method. Practical considerations about its applicability and asymptotic behavior analysis for theoretical justification are also provided.