Multi-metrics Classification Machine

Distance and its related decision rules are important in classification problems. kNN classifies a data point by the labels of its k-nearest neighbors and can be ameliorated by metric learning. For SVM, representative hyperplanes are found to refer to the location of every class and any point can be labeled by its perpendicular distance from the hyperplanes. Inspired by metric learning and SVM, a multi-metric classification machine, called MMCM, with a new prediction mechanism is proposed based on a novel distance relationship discerned by multi-metrics learning of the specificity information of each class. MMCM aims to find multi-metrics, namely the multiple local linear transformations for each class, to map data points into a new feature space, in which the distance between a point and its corresponding class centroid is minimized and data points of other classes are far from the centroid. An example with unknown label is classified according to the label of its nearest centroid. The primary problem is slacked as a linear optimization problem and kernel is introduced to make a nonlinear transformation. Enormous experiments verify MMCM's competitive performance both on binary classification and multi-class classification compared to state-of-the-art classification methods.