2
$\begingroup$

In the follow-up to this Ways to understand 2-dimensional time-series data

I'm working on 2D time series data where two attributes are depth and temperature. When I plotted depth-vs-temp curve and saw its variation over time, the fluctuation occurs at few places only.

Lets say temperature is dependent on depth and it varies at few depths over time.

Would doing cluster analysis on each time sample and comparing it with the next one be good? The point is to statistically identify changes across timestamps. And then decide which are significant timestamp samples.

What models/papers should I study to get insights out of the data?

Improve this question
$\endgroup$
7
  • 1
    $\begingroup$ Any chance you could upload the plot? I'm not sure that I'm sure I know what you're describing. Also, what else do you know about the data? How was it collected, how many observations, etc. $\endgroup$
    – shadowtalker
    Commented Jun 22, 2014 at 9:59
  • $\begingroup$ Temperature samples are recorded for 1k depth samples. And this is recorded over 500 time samples, say 1-2 minutes apart. Depth samples are uniformaly distributed over a range. But, due to decimal corrections, over time samples depth values needn't be same. Makes it clear? Sorry for being cryptic. $\endgroup$
    – piroot
    Commented Jun 22, 2014 at 10:04
  • $\begingroup$ That helps. I'd still like to see the plot, or some kind of mock-up if the data is confidential. $\endgroup$
    – shadowtalker
    Commented Jun 22, 2014 at 10:04
  • $\begingroup$ If you could tell me what inferences you would like to know. I'd prepare a mock-up based on that. $\endgroup$
    – piroot
    Commented Jun 22, 2014 at 10:10
  • $\begingroup$ "when I plotted depth-vs-temp curve and saw its variation over time, the fluctuation occurs at few places only." $\endgroup$
    – shadowtalker
    Commented Jun 22, 2014 at 10:13

1 Answer 1

Reset to default
2
$\begingroup$

Based on your comments, my immediate instinct is to fit a regression line at each time point, and then plot the coefficient(s) over time.

My next instinct is to compute a "distance" between each depth band using one of several distance functions. Then you can run cluster analysis on it directly, e.g. single-linkage cluster analysis. You could also apply multidimensional scaling, principal components analysis, or a network/graph representation and optionally run analyses on those.

There's an overview of the subject here (pdf slides). Papers reviewing the subject here and here.

Improve this answer
$\endgroup$
1
  • $\begingroup$ I have just implemented both the ideas. Thanks. I'd try and further dive into pca and networks approach. $\endgroup$
    – piroot
    Commented Jun 22, 2014 at 18:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.