Clustering Time Series with DTW

Hi everyone.

I have ~161 time series of heart rates taken during a vocalization. I would like to sort these using the DTW algorithm. I have tried using the following to do this:

[idx,c,sumd,d] = kmedoids(dat,nclust,'Distance',@dtw);

But I end up with the following errors. I have done this before... with the exact same code. Does anyone know what I could be doing wrong?

Error using pdist (line 371)
Error evaluating distance function 'dtw'.

Error in internal.stats.kmedoidsDistObj/pdist (line 65)
                out = pdist(X,distObj.distance);

Error in kmedoids>precalcDistance (line 575)
distVec = distObj.pdist(X);

Error in kmedoids>loopBody (line 546)
            xDist = precalcDistance(X,distObj);

Error in kmedoids (line 362)
oneRun = loopBody(algorithm,initialize,X,k,distObj,pneighbors,numsamples,options,display,usePool,onlinePhase);

Error in dtwclassifier (line 9)
[idx,c,sumd,d] = kmedoids(dat,nclust,'Distance',@dtw);

Caused by:
    Error using dtw (line 115)
    The number of rows between X and Y must be equal when X and Y are matrices

Would it be correct to compute my own distance matrix using DTW, and perform K-Means clustering on that?

Any help would be absolutely amazing! Thank you!

Answer 1

[Srivardhan Gadila]

I think the error is due the reason that dtw function operates on 2 signals only and the output is always a scalar.

As per the description of 'Distance' Name Value Pair Argument of the kmedoids function:

"See pdist for the definition of each distance metric. kmedoids supports all distance metrics supported by pdist."

And as per the description of Distance input argument of the pdist function for custom distance:

"@distfun

Custom distance function handle. A distance function has the form

function D2 = distfun(ZI,ZJ)
% calculation of distance
...

where

• ZI is a 1-by-n vector containing a single observation.
• ZJ is an m2-by-n matrix containing multiple observations. distfun must accept a matrix ZJ with an arbitrary number of observations.
• D2 is an m2-by-1 vector of distances, and D2(k) is the distance between observations ZI and ZJ(k,:)."

Hence you can't use the dtw function handle directly and you can use it as follows:

data = rand(161,20);

[idx,c,sumd,d] = kmedoids(data,10,'Distance',@dtwf);


function dist = dtwf(x,y)
% n = numel(x);
m2 = size(y,1);

dist = zeros(m2,1);

for i=1:m2
    dist(i) = dtw(x,y(i,:));
end
end

[Manash Sahoo]

This is fantastic! Thank you so much.

[Amila]

Dear Srivardhan Gadila,

Thank you for giving helping, this was helped mee too

I'm trying to use same kind of test the problem is my time series (raw) are not same length i alined them with a time and make a matrics somthing like data = rand(161,20); but the problem is due to un even size (time), some raws has NaN at the 'end' and the 'beginning;

could you kindly helped me ...

my dat is look like this

NaN NaN NaN 0.5652 ... 0.9338 NaN NaN NaN

NaN NaN 0.6514 0.6486 ... 0.8719 NaN NaN NaN

NaN NaN 0.4987 0.7981 ... 0.3011 0.8229 NaN NaN

NaN NaN 0.2845 0.2204 ... 0.2360 0.6886 0.8708 NaN

NaN NaN 0.8306 0.8579 ... 0.8316 0.6039 0.3528 0.7837

NaN 0.1909 0.8184 0.9047 ... 0.4378 0.3870 0.4002 0.6733

0.7114 0.4286 0.9382 0.2920 ... 0.5071 0.0655 0.5979 0.8407

0.7834 0.0145 3.2610e-04 0.7259 ... 0.1855 0.9986 0.9115 0.4209

0.6239 0.3253 0.6404 0.3394 ... 0.1515 0.6610 0.1330 0.7037

... ... ... ... ... ... ... ... ... ...

When I run the codes all the raws that have NaN value going to omit how i can avoid this

please kindly help me

Answer 2

[Amila]

If anyone has any idea on how to do it please give me some advice !

Thank you