An Experimental Study on Neighbourhood Component Analysis

In this experimentation project, we introduce the Neighbourhood Component Analysis (NCA), a classification method combining k-nearest neighbour (KNN) and learned distance metric, originally proposed by Goldberger et al. in 2004. With an implementation in the R environment, we perform experiments on several datasets. Results of our experiments suggest that with the learned distance metric, KNN classification can be improved, especially for skewed and noisy data. For certain type of data, NCA also outperforms traditional dimension reduction methods. However, the comparisons to other classifiers, i.e. linear and quadratic discriminant analysis, do not show significant strength in this method.