2
$\begingroup$

My dataset:

  • 121 individuals characterized by a categorical variable let’s say Y (so 121 values of Y {low, medium, high})

  • for each individual, I have 5 time series, let’s say $X_1$, $X_2$, $X_3$, $X_4$, $X_5$.

  • each time series contains 6000 points.

All the time series increase over time, with abrupt changes, and end with a plateau.

I want to build a classification model to understand which time series (or features of the time series, like abrupt changes, or combination of these features) explain and then predict the best the variable Y.

I am not used to work with time series, and I have found that the "Time series" analysis models are mostly around forecasting from existing data, and so I spent time looking for a clear procedure explaining how to handle this type of data for this specific purpose without much success.

What I have tested (see below) is always based on one time serie (matrix of 121 * 6000) at once since I do not understand how to deal with several time series:

  • unsupervised clustering of each time serie (hclust and kmeans): in order to see if we retrieve the groups of Y {low, medium, large}
  • a Discrete Wavelet Transform (DWT) to extract features and application of a decision tree on the extracted features. Here I do not understand how to interpret the features (wavelet coefficients and scaling coefficient) regarding to the initial data.

Please see the image below to understand data

enter image description here

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.