Enhancing structure learning of Markov network using Alternating Decision Trees

Most of the existing algorithms used for the purpose of Markov network structure learning are either restricted to learning interactions among small number of variables or are extremely slow in nature because of the large number of possible structures, as given in the literature. In this paper, we propose a novel method of using Alternating Decision (AD) trees for learning Markov network structures. The advantage in using the AD trees is that complex interactions among many variables can be represented and high prediction accuracy is obtained. Given a data set, using AD trees for structure learning involves learning an AD tree for the prediction of each variable, conversion of each tree to a set of conjunctive features and weight learning. The set of conjunctive features define the Markov network structure. In this paper, we compare Decision Tree Structure Learning (DTSL) method proposed by Lowd and Davis with the AD tree approach empirically over five data sets and find AD tree approach to be more accurate than the DTSL approach for two data sets. However, the method using AD trees is slower than the DTSL method.