Dependency Tree Hidden Markov Models
Joakim Jiten, Bernard Merialdo

Project description

We are studying a new type of multi-dimensional Hidden Markov Model. Multidimensional HMMs have an exponential complexity which makes them unusable in practise. The main idea of DT-HMM is to replace the multiple dependency of the neighbours (2 neighbours in the case of 2 dimensions) with a single dependency with a randomly selected neighbour. Multiple dependencies can be weakly taken into account by the use of several dependency trees.

We have developped a library to manipulate DT-HMMs and applied them to a number of different tasks, such as: