Incremental Construction of Structured Hidden Markov Models

Ugo Galassi, Attilio Giordana, Lorenza Saitta

This paper presents an algorithm for inferring a Structured Hidden Markov Model (S-HMM) from a set of sequences. The S-HMMs are a sub-class of the Hierarchical Hidden Markov Models and are well suited to problems of process/user profiling. The learning algorithm is unsupervised, and follows a mixed bottom-up/top-down strategy, in which elementary facts in the sequences (motifs) are progressively grouped, thus building up the abstraction hierarchy of a S-HMM, layer after layer. The algorithm is validated on a suite of artificial datasets, where the challenge for the learning algorithm is to reconstruct the model that generated the data. Then, an application to a real problem of molecular biology is briefly described.