||Nearest-neighbor approaches for classification have long been recognized for their potential in achieving low error rates, despite their perceived lack of scalability. Recent advances in the efficient computation of approximate k-nearest neighborhoods have made the nearest neighbor approaches more affordable in practice. However, their effectiveness is still limited due to their sensitivity to noise and to the choice of neighborhood size k. In this paper, we propose a general-purpose method for nearest-neighbor classification that seeks to compensate for the effects of noise through the determination of natural clusters in the vicinity of the test item. The classification model, based on elements of the relevant-set correlation (RSC) model for clustering, also allows for the automatic determination of an appropriate value of k for each test item. We also provide experimental results that demonstrate the competitiveness of our approach with that of other popular classification methods.