target vector generation

[arashtavakoli]

Hi,

I am using TSFRESH for retreiving features from sensor data, similar to the robotic example. In summary, I have a dataframe say with 10000 rows and for each row I have a label (e.g., row 1: eating, row 2: drinking, etc). I know that I can run feature extraction on the numeric parts but how to retrieve label (target) vector for these features? Does TSFRESH automatically do this? Note that I am trying to run feature extraction with 50 percent overlapping and the final feature extraction matrix won't be the same as the starting data frame thus the feature extracted data frame would be smaller. I am basically confused to how to map these features to the original target vector that I had.

[arashtavakoli]

kindly, will there be any response to this?

[nils-braun]

Sorry @arashtavakoli for the late answer and thanks for pushing the issue again. tsfresh always works in time series. So if you would like to predict a label for each row just using the data of that row, this would not work. However, I assume you would like to use the data up to this row (e.g. to predict the label for row 123, maybe use the sensor measurements from 100-122). Here tsfresh can help. We call this "rolling" and our documentation might give you a lot of information. In short: you cut your data into windows, each let's say with the size of 40 and then produce features on each of those. In this case, you can just use the target vector that you already have (except maybe at the borders, where the windows are too small).

Hope that helped? Maybe that is already what you were thinking of, in this case a small example of your problem would help :-)