Hierarchical Multilabel Classification with Optimal Path Prediction

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