Hierarchical Multilabel Classification with Optimal Path Prediction

We consider multilabel classification problems where the labels are arranged hierarchically in a tree or directed acyclic graph (DAG). In this context, it is of much interest to select a well-connected subset of nodes which best preserve the label dependencies according to the learned models. Top-down or bottom-up procedures for labelling the nodes in the hierarchy have recently been proposed, but they rely largely on pairwise interactions, thus susceptible to get stuck in local optima. In this paper, we remedy this problem by directly finding a small number of label paths that can cover the desired subgraph in a tree/DAG. To estimate the high-dimensional label vector, we adopt the advantages of partial least squares techniques which perform simultaneous projections of the feature and label space, while constructing sound linear models between them. We then show that the optimal label prediction problem with hierarchy constraints can be reasonably transformed into the optimal path prediction problem with the structured sparsity penalties. The introduction of path selection models further allows us to leverage the efficient network flow solvers with polynomial time complexity. The experimental results validate the promising performance of the proposed algorithm in comparison to the state-of-the-art algorithms on both tree- and DAG-structured data sets.