Learning approximate MRFs from large transaction data

Abstract

In this paper we consider the problem of learning approximate Markov Random Fields (MRFs) from large transaction data. We rely on frequent itemsets to learn MRFs on the data. Since learning exact large MRFs is generally intractable, we resort to learning approximate MRFs. Our proposed modeling approach first employs graph partitioning to cluster variables into balanced disjoint partitions, and then augments important interactions across partitions to capture interdependencies across them. A novel treewidth based augmentation scheme is proposed to boost performance. We learn an exact local MRF for each partition and then combine all the local MRFs together to derive a global model of the data. A greedy approximate inference scheme is developed on this global model. We demonstrate the use of the learned MRFs on the selectivity estimation problem. Empirical evaluation on real datasets demonstrates the advantage of our approach over extant solutions. © Springer-Verlag Berlin Heidelberg 2006.