Clustering HMM States for Symbolic Description of Tempo- ral Sequences

We seek a high level abstraction not directly on states level, but as a function over states. By constructing the function we identify substructures in the topology of HMM that correspond to “macro” status. In order to construct the function, additional information – sparse manually labeled or automatically generated sources – is needed and task dependent. Different functions can be constructed for different tasks. On the other hand, from the resulting functions one can tell whether a task-dependent structure is captured by the HMM learned from historic data.