On Maximum Margin Hierarchical Multilabel Classification

We present work in progress towards maximum margin hierarchical classification where the objects are allowed to belong to more than one category at a time. The classification hierarchy is represented as a Markov network equipped with an exponential family defined on the edges. We present a variation of the maximum margin multilabel learning framework, suited to the hierarchical classification task and allows efficient implementation via gradient-based methods. We compare the behaviour of the proposed method to the recently introduced hierarchical regularized least squares classifier as well as two SVM variants in Reuter’s news article classification. Often in hierarchical classification, the object to be classified is assumed to belong to exactly one (leaf) node in the hierarchy (c.f. [5, 2, 4]). Following [3], in this paper we consider the more general case where a single object can be classified into several categories in the hierarchy, to be specific, the multilabel is a union of partial paths in the hierarchy. For example, a news article about David and Victoria Beckham could belong to partial paths sport, football and entertainment, music but might not belong to any leaf categories such as champions league or jazz. In our setting the training data ((xi,y(xi))) m i=1 consists of pairs (x,y) of vector x ∈ R and a multilabel y ∈ {+1,−1} consisting of k microlabels. As the model class we use the exponential family