Short sequences

[denklewer]

Hello,
Is it possible to use your library to train HMM for short sequences?
I have some short sequeces (vectors of length from 8 to 25) And want to build the model which will be able to return probability score for each sequence.
All your examples contain 1 big sequence. So I'm wondering if it is possible to use multiple for training?

[lorek]

We have not implemented training on multiple sequences. Note that we have two different training methods:
a) F^Q -- which relies on co-occurrence matrix
b) F^ML -- which optimises the log-likelihood of observations.
Implementing training on multiple sequences in a) is quite straightforward (it is no effort to compute the co-occurrence matrix then). However, implementing it in b) would require more work (but is doable). Mainly because, by default we sample shorter subseqences.

Of course you can always concatenate your sequences, be since there are short and you have plenty of them -- however a distortion on "connections" could influence everything significantly.