Improving performance of nearest neighborhood classifier using genetic programming
Abdul Majid, Asifullah Khan, Anwar Majeed Mirza · 2005
Nearest neighborhood classifier (kNN) is most widely used in pattern recognition applications. Depending on the selection of voting methodology, the problem of outliers has been encountered in this classifier. Therefore, selection and optimization of the voting methodology is very important. In this work, we have used Genetic Programming (GP) to improve the performance of nearest neighbor classifier. Instead of using predefined k nearest neighbors, the number of men and women in the first two quartiles in Euclidean space are used for voting. GP is, then, used to evolve an optimal class mapping function that effectively reduces the outliers. The performance of modified nearest neighborhood (ModNN) classifier is then compared with the conventional kNN for gender classification problem. Receiver Operating Characteristics curve and its Area Under the Convex Hull (A UCH) are used as the performance measures. Considering the first three and first five eigen features respectively, ModNN achieves AUCH equal to 0.985 and 0.992 as compared to 0.9693 and 0.9795 of conventional kNN respectively.