A General Probabilistic Framework for Mining Labeled Ordered Trees

Abstract

We propose a new probabilistic model for mining labeled ordered trees. A noteworthy feature of the proposed model is to consider ordered siblings by modeling the dependencies of a node in a tree on the elder sibling as well as the parent. This model is reasonably extended from a variety of existing probabilistic models for strings and trees. We further propose a new learning/mining method to estimate the parameters of this model, based on an EM algorithm. This is also an extension of those for various simpler probabilistic models, such as hidden Markov models and hidden tree Markov models. We evaluated the effectiveness of our proposed method using both synthetic and real-world data sets, comparing the results with those of several simpler probabilistic models. Experimental results have shown that our proposed method outperforms the other methods compared, being statistically significant in all cases tested. This result tells us that the proposed methodology is highly effective for mining labeled ordered trees, which have recently emerged as one of the typical data structures in numerous data mining domains, including the web, text mining and bioinformatics.