Learning Structural Variations in Shock Trees

Andrea Torsello, Edwin R. Hancock

Abstract: In this paper we investigate how to construct a shape space for sets of shock trees. To do this we construct a super-tree to span the union of the set of shock trees. We learn this super-tree and the correspondences of the node in the sample trees using a maximizing likelihood approach. We show that the likelihood is maximized by the set of correspondences that minimizes the sum of the tree edit distance between pair of trees, subject to edge consistency constraints. Each node of the super-tree corresponds to a dimension of the pattern space. Individual such trees are mapped to vectors in this pattern space.