In my final year project I'm using HMM to predict children depression status as I found that there are patterns which can be identified by their activities.
I have 3 hidden states, Depressed, Non-Depressed, Clinical Range
For observation sequence I have child daily activities.
So I need to find hidden state for a given observation sequence. For that I'm using Viterbi Algorithm.
So example of my observation sequence is which is input for the Viterbi algorythm is:
sleep,sleep,sleep,sleep,cry,cry,cry,cry,sleep,sleep,wake,wake,sleep,sleep,cry,eat,eat,eat,fight,fight,play,play,play,play,play,fight,cry,cry.....................
Then the output that I get is that most likely series of events was {depressed,depressed,depressed,clinical,clinical,. . . . . . . . . . . }
I'm not sure if I'm getting the right output from Viterbi algorithm
Is my observation sequence correct. Having like 50 observations in the observations sequence and 3 hidden state?
How did you derive your transmission and emission probabilities? Did you train the HMM on a training set?
Yes, I trained the HMM and get transition and emission probabilities.
If you have the transition and emission probability matrices available, then you can test the HMM on a smaller set and precompute the optimal path by hand.
That should show you if the hidden states you calculated are the same ones as proposed by the HMM on the smaller set.
Whether or not your predicted sequence of hidden states is correct depends also on the size of your training set. And sometimes also upon the choice of initial parameters. But I cannot answer you if your results are correct for your particular dataset. Remember that the Viterbi path is probabilistically the most likely sequence of hidden states, according to your observations, but it is not necessarily also the true one.