A new pattern classification improvement method with local quality matrix based on K-NN

In order to pursue the high accuracy in classification task, we propose a new pattern classification accuracy improvement (CAI) method working with local quality matrix. The quality matrix expresses the conditional probability of the object belonging to one class when classified to another class, and it is estimated based on the K-nearest neighbors (K-NN) of object. The K-NN selected from the labeled data set are classified using a given classifier, and their classification results represented by either probability or belief functions could be modified using a quality matrix to make them as close as possible to the ground truth. Hence, the quality matrix can be estimated by minimizing the sum of distances between the modified classification results of the K-NN and the truth. In the optimization procedure, the K-NN are considered with different weights depending on their distances to the object. The smaller distance, the bigger weight. This is to control the influence of the neighbors far from object. The classification result of each object will be corrected by the corresponding optimized quality matrix via an appropriate probability redistribution way. The performance of CAI method is tested with respect to other related methods using some real data sets. The experimental results show that CAI substantially improves the classification accuracy, and it is robust to the choice of K value.