MC: An Unsupervised Data Preprocessing for Classification

The generalization ability of a classifier is often inherently associated with both the intra-class compactness and the inter-class separability. Owing to the fact that some current lower-dimensional manifold embedding techniques as a preprocessing for classification learning often lead to poor performance, in this paper, a new unsupervised data preprocessing technique called as manifold contraction (MC) is proposed for the subsequent classification task. The main contribution of our MC lies in: 1) the intra-manifold scatter becomes smaller while the inter-manifold scatter gets bigger relatively by a proper contraction mapping; 2) different from dimensionality reduction techniques, the estimation of intrinsic dimensionality can be avoided. The final experimental results show that MC preprocessing technique is effective and promising in the subsequent classification task especially in small-size labeled samples case.