Mining Predictive Redescriptions with Trees

In many areas of science, scientists need to find distinct common characterizations of the same objects and, vice versa, identify sets of objects that admit multiple shared descriptions. For example, a biologist might want to find a set of bioclimatic conditions and a set of species, such that this bioclimatic profile adequately characterizes the areas inhabited by these fauna. In data analysis, the task of automatically generating such alternative characterizations is called redescription mining. A number of algorithms have been proposed for mining redescriptions which usually differ on the type of redescriptions they construct. In this paper, we demonstrate the power of tree-based redescriptions and present two new algorithms for mining them. Tree-based redescriptions can have very strong predictive power (i.e. they generalize well to unseen data), but unfortunately they are not always easy to interpret. To alleviate this major drawback, we present an adapted visualization, integrated into an existing interactive mining framework.