Complexity of Computing Distances between Geometric Trees [chapter]

Aasa Feragen
2012 Lecture Notes in Computer Science  
Geometric trees can be formalized as unordered combinatorial trees whose edges are endowed with geometric information. Examples are skeleta of shapes from images; anatomical tree-structures such as blood vessels; or phylogenetic trees. An inter-tree distance measure is a basic prerequisite for many pattern recognition and machine learning methods to work on anatomical, phylogenetic or skeletal trees. Standard distance measures between trees, such as tree edit distance, can be readily translated
more » ... to the geometric tree setting. It is well-known that the tree edit distance for unordered trees is generally NP complete to compute. However, the classical proof of NP completeness depends on a particular case of edit distance with integer edit costs for trees with discrete labels, and does not obviously carry over to the class of geometric trees. The reason is that edge geometry is encoded in continuous scalar or vector attributes, allowing for continuous edit paths from one tree to another, rather than finite, discrete edit sequences with discrete costs for discrete label sets. In this paper, we explain why the proof does not carry over directly to the continuous setting, and why it does not work for the important class of trees with scalar-valued edge attributes, such as edge length. We prove the NP completeness of tree edit distance and another natural distance measure, QED, for geometric trees with vector valued edge attributes.
doi:10.1007/978-3-642-34166-3_10 fatcat:myge5amuajd7zdnwvlzhwfkptm