A new evidential K-nearest neighbor rule based on contextual discounting with partially supervised learning

Abstract

The evidential K nearest neighbor classifier is based on discounting evidence from learning instances in a neighborhood of the pattern to be classified. To adapt the method to partially supervised data, we propose to replace the classical discounting operation by contextual discounting, a more complex operation based on as many discount rates as classes. The parameters of the method are tuned by maximizing the evidential likelihood, an extension of the likelihood function based on uncertain data. The resulting classifier is shown to outperform alternative methods in partially supervised learning tasks.