The nearest neighbor algorithm is one of the most classical pattern recognition algorithms, which classification performance highly depends on the distance metric between samples. Appropriate distance metric can help improve the classification performance of the algorithm. However, such algorithms mostly seek metrics to maximize classification effectiveness from the perspective of discriminant models currently, ignoring the aggregation properties of various sample sets belonging to different classes. In view of this, a nearest neighbor classification algorithm based on minimum component eigenvector subspace projection (NN_MCESP) is proposed. This algorithm combines classic principal component analysis (PCA) and nearest neighbor algorithm, which can effectively implement aggregation properties analysis of various sample clusters based on minimum component eigenvector projection, and complete classification based on subspace nearest neighbor voting criteria. The effectiveness and stability of performance NN_MCESP are validated by comparing with other classification algorithms on multiple data sets.