Finding low-entropy sets and trees from binary data
The discovery of subsets with special properties from binary data has been one of the key themes in pattern discovery. Pattern classes such as frequent itemsets stress the co-occurrence of the value 1 in the data. While this choice makes sense in the context of sparse binary data, it disregards potentially interesting subsets of attributes that have some other type of dependency structure. We consider the problem of finding all subsets of attributes that have low complexity. The complexity is measured by either the entropy of the projection of the data on the subset, or the entropy of the data for the subset when modeled using a Bayesian tree, with downward or upward pointing edges. We show that the entropy measure on sets has a monotonicity property, and thus a levelwise approach can find all low-entropy itemsets. We also show that the treebased measures are bounded above by the entropy of the corresponding itemset, allowing similar algorithms to be used for finding low-entropy trees. We describe algorithms for finding all subsets satisfying an entropy condition. We give an extensive empirical evaluation of the performance of the methods both on synthetic and on real data. We also discuss the search for high-entropy subsets and the computation of the Vapnik-Chervonenkis dimension of the data.
Author: Hannes Heikinheimo, Helsinki University of Technology