Exclusive lasso-based k-nearest-neighbor classification

Conventionally, the k nearest-neighbor (kNN) classification is implemented with the use of the Euclidean distance-based measures, which are mainly the one-to-one similarity relationships such as to lose the connections between different samples. As a strategy to alleviate this issue, the coefficients coded by sparse representation have played a role of similarity gauger for nearest-neighbor classification as well. Although SR coefficients enjoy remarkable discrimination nature as a one-to-many relationship, it carries out variable selection at the individual level so that possible inherent group structure is ignored. In order to make the most of information implied in the group structure, this paper employs the exclusive lasso strategy to perform the similarity evaluation in two novel nearest-neighbor classification methods. Experimental results on both benchmark data sets and the face recognition problem demonstrate that the EL-based kNN method outperforms certain state-of-the-art classification techniques and existing representation-based nearest-neighbor approaches, in terms of both the size of feature reduction and the classification accuracy.