On Thu, Oct 18, 2012 at 12:18 AM, Didier Vila <[email protected]> wrote:
> Should I maximize the “score” function depending the number of state
> through several candidates that makes senses from a business point of view
> ?
>
The best criterion to optimize is not "score" but the evaluation metric you
really care about for your application. For example, if what you want to do
is time-series classification, you should optimize the number of states for
classification accuracy. Also, be careful, you cannot use your test data to
choose the number of states. You need to use validation data or
cross-validation.
For some types of data, states have an intuitive meaning and the number of
states can be chosen easily. For example, if what you want to do is mouse
gesture recognition, you can use one state for your HMM modeling a straight
line and two states for your HMM modeling a gesture with a turn. You can
also try to visualize your data and see if it has an easy-to-identify
number of "modes".
In an old paper of mine [*], I proposed a simple heuristic to choose the
number of states but it's only useful if you want to model several object
classes, each with one HMM. In that case, choosing the optimal number of
states for all HMMs is a combinatorial problem and my heuristic allows to
search only 1 parameter.
You may also want to try GMMHMM but in that case you need to tune the
number of mixture components as well (more components give your HMM more
expressive power therefore your HMM may need fewer states).
HTH,
Mathieu
[*] http://mblondel.org/publications/mblondel-icpr2010.pdf
------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_sfd2d_oct
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
