Hierarchical Classification via Orthogonal Transfer

摘要: We consider multiclass classification problems where the set of labels are organized hierarchically as a category tree. associate each node in tree with classifier and classify examples recursively from root to leaves. propose hierarchical Support Vector Machine (SVM) that encourages at be different classifiers its ancestors. More specifically, we introduce regularizations force normal vector classifying hyperplane orthogonal those ancestors much possible. establish conditions under which training such SVM is convex optimization problem, develop an efficient dual-averaging method for solving it.