Detecting irrelevant subtrees to improve probabilistic learning from tree-structured data

In front of the large increase of the available amount of structured data (such as XML documents), many algorithms have emerged for dealing with tree-structured data. In this article, we present a probabilistic approach which aims at a priori pruning noisy or irrelevant subtrees in a set of trees. The originality of this approach, in comparison with classic data reduction techniques, comes from the fact that only a part of a tree (i.e. a subtree) can be deleted, rather than the whole tree itself. Our method is based on the use of confidence intervals, on a partition of subtrees, computed according to a given probability distribution. We propose an original approach to assess these intervals on tree-structured data and we experimentally show its interest in the presence of noise.