Pattern Synthesis Using Fuzzy Partitions of the Feature Set for Nearest Neighbor Classifier Design

Nearest neighbor classifiers require a larger training set in order to achieve a better classification accuracy. For a higher dimensional data, if the training set size is small, it suffers from the curse of dimensionality effect and performance gets degraded. Partition based pattern synthesis is an existing technique of generating a larger set of artificial training patterns based on a chosen partition of the feature set. If the blocks of the partition are statistically independent then the quality of synthetic patterns generated is high. But, such a partition, often does not exist for real world problems. So, approximate ways of generating a partition based on correlation coefficient values between pairs of features were used earlier in some studies. That is, an approximate hard partition, where each feature belongs to exactly one cluster (block) of the partition was used for doing the synthesis. The current paper proposes an improvement over this. Instead of having a hard approximate partition, a soft approximate partition based on fuzzy set theory could be beneficial. The present paper proposes such a fuzzy partitioning method of the feature set called fuzzy partition around medoids (fuzzy-PAM). Experimentally, using some standard data-sets, it is demonstrated that the fuzzy partition based synthetic patters are better as for as the classification accuracy is concerned.