Metric learning by discriminant neighborhood embedding

In this paper, we learn a distance metric in favor of classification task from available labeled samples. Multi-class data points are supposed to be pulled or pushed by discriminant neighbors. We define a discriminant adjacent matrix in favor of classification task and learn a map transforming input data into a new space such that intra-class neighbors become even more nearby while extra-class neighbors become as far away from each other as possible. Our method is non-parametric, non-iterative, and immune to small sample size (SSS) problem. Target dimensionality of the new space is selected by spectral analysis in the proposed method. Experiments on real-world data sets demonstrate the effectiveness of our method.