The slides illustrates a projected Newton method to optimize a multilabel structured output learning model with random spanning tree approximation.