Random subspace and random projection nearest neighbor ensembles for high dimensional data

The random subspace and the projection methods are investigated compared as techniques for forming ensembles of nearest neighbor classifiers in high dimensional feature spaces. two have been empirically evaluated on three types high-dimensional datasets: microarrays, chemoinformatics, images. Experimental results 34 datasets show that both method lead to improvements predictive performance using standard classifier, while best use depends type data considered; microarray chemoinformatics datasets, outperforms method, opposite holds image datasets. An analysis complexity measures, such attribute instance ratio Fisher’s discriminant ratio, provide some more detailed indications what relative can be expected specific also indicate resulting may competitive with state-of-the-art ensemble classifiers; perform par forests