Relationship between Variants of One-Class Nearest Neighbours and Creating their Accurate Ensembles

In one-class classification problems, only the data for the target class is available, whereas the data for the non-target class may be completely absent. In this paper, we study one-class nearest neighbour (OCNN) classifiers and their different variants. We present a theoretical analysis to show the relationships among different variants of OCNN that may use different neighbours or thresholds to identify unseen examples of the non-target class. We also present a method based on inter-quartile range for optimizing parameters used in OCNN in the absence of non-target data during training. Then, we propose to use two ensemble approaches based on random subspace and random projection methods to create accurate OCNN ensembles that significantly outperforms the baseline OCNN. We tested the proposed methods on 15 benchmark and real world domain-specific datasets to show their superior performance. The results give strong evidence that the random projection ensemble of the proposed OCNN with optimized parameters perform significantly and consistently better than the single OCNN on all the tested datasets.