A new fuzzy k-Nearest Neighbors rule in pattern recognition
Résumé
A new fuzzy k-nearest neighbours (k-NN) rule is proposed in this article. On considering four feature variables in a k-NN methodology, a fuzzy class membership function is constructed. These four variables are the number of prototypes of same class in k-Nearest Neighbours of an incoming pattern, the distances from this incoming pattern to its k-Nearest Neighbours, the rank level of k-Nearest Neighbours, and finally, the variation in distances among the k-Nearest Neighbours. Three application examples are used to test our proposed rule. The number of samples misclassified or successful classification rate is calculated and presented in the results section. The fuzzy k-NN rule proposed seems to be very flexible and is adaptable according to the nature of the application. It yields good and satisfactory results when compared with other rules used in this domain, provided that the parameter, which is introduced to assign fuzzy weighting to each neighbour of an incoming pattern according to the importance of the evidence carried by that neighbour, is firstly well optimised.