Estimate Markov Chain Transition Matrix in MATLAB With Different State Sequence Lengths

Question

I'm trying to build the transition matrix for a Markov Chain in MATLAB; I have several different observation sequences (all of varying lengths) and I need to generate the transition matrix using those.

Constructing a multi-order Markov chain transition matrix in Matlab shows me how to build a transition matrix with a single observation sequence.

How can I construct one using observations of different length? One example can be that one sequence is 1,2,3,4 and another is 4,5,6. Is there any way to do this without having to for loop through all sequences and computing counts?

Answer 1

So for Markov chains, I assume you're only interested in the state transitions. You could group all state transitions into a single Nx2 matrix and then count the number of times a row appears.

For this example I'm using three observations of length 4, 3, and 3. I can use cellfun to group all the state transitions together in a single matrix in the following way:

obs = cell(1, 3);

obs(1) = {[1 2 3 4]};
obs(2) = {[4 5 6]};
obs(3) = {[3 4 5]};

transitions = cellfun(@(x)([x(1:length(x)-1); x(2:length(x))]), obs, 'UniformOutput', false);

alltransitions = cell2mat(transitions)';

Which gives me my observed transitions (1->2, 2->3, 3->4 ...):

alltransitions =

     1     2
     2     3
     3     4
     4     5
     5     6
     3     4
     4     5

To set up the transition matrix, you could take the advice listed here, and count the rows of all of your transitions:

http://www.mathworks.it/matlabcentral/answers/75009-i-ve-a-matrix-of-6x4-and-i-want-to-count-the-rows-how-many-times-it-occur-in-a-matrix

[uniqueTransitions, ~, i]=unique(alltransitions,'rows','stable');
v=arrayfun(@(x) sum(i==x),1:size(uniqueTransitions,1))';
p = v/sum(v);

My vector p contains my transition probability, so I can then go ahead and construct a sparse matrix

transitionMatrix = sparse(uniqueTransitions(:,1), uniqueTransitions(:,2), p, 6,6)

which results in:

transitionMatrix =

   (1,2)       0.1429
   (2,3)       0.1429
   (3,4)       0.2857
   (4,5)       0.2857
   (5,6)       0.1429

Answer 2

This maybe an easier way for calculating the transition probability matrix (TPM) for a given sequence of data (one vector) like this:

myS = {S1,S2,S1,S3,...} with as many states as you have;

TPM = hmmestimate(myS,myS);

hmmestimate function is defined under hidden markov models in MATLAB.