-1

my task is to classify an observation vector (features taken from sensors signals) to one of M states. I can use multiclass classifiers (trees or SVM) but since these states are dependent i am looking for a model to use the dependency between the states. I thought of using an HHM (i am new to HMM) but from my research i mostly found the emission model to require a complete symbols set with probabilities or some generative distribution model. Both seem unfeasible.

  1. How can I use the discriminative tools i know (trees , SVM) to train a suitable HMM ?
  2. If it is not an appropriate model for my problem which one is ?

Note: I am Classifying human physical activity from an on-body Accelerometer. for example i may want to classify between these states: 'standing still' ' walking' & 'running'. Instead of just classifying each frame by itself i would like to treat this as a sequence. this way, for example, if i have an observation that looks like 'running' , the next observation is most likely to be 'running' too and not confused with walking which looks similar and possible but less likely.

user3730223
  • 1
  • Search specifically for HSMM (hidden Semi-Markov Models) because this variation takes into account sequences as a whole – rpd Dec 08 '15 at 14:07
  • thanks, but i am less bothered by the states transition and more by the observations issue. also any idea why i got negative rating, am i missing something obvious here ? – user3730223 Dec 09 '15 at 08:22
  • The negative rating goes to the fact that the question seems rather vague and not directly related to programming (which is actually true) but I happen to have experience in the topic so I decided to comment. As for your concern, have you taken a look at this?: http://stackoverflow.com/questions/17487356/hidden-markov-model-for-multiple-observed-variables – rpd Dec 09 '15 at 08:41
  • thanks , great link , CRF might be what i need. – user3730223 Dec 14 '15 at 20:21

1 Answers1

0

to my understanding, HMM must use generative emission model. A hybrid approach can use discriminative methods to build emission probabilities as in combined hmm & svm.

user3730223
  • 1