Hi,
I have a multivariate time series (3D) dataset of the shape, (samples, sequences, features). For instance, (patients, time_sequences_in_hour, features), whereby each sample is a distinct group of time sequences.
The feature selection libraries in Python that I have surveyed so far are on tabular (2D) data, i.e: ReliefF, https://epistasislab.github.io/scikit-rebate/. Is there a feature selection library in Python that works well for multivariate time series dataset as I mentioned above?
Thank you.
Based on the shape, sounds like input to RNNs, specially LSTMs. Then you can perturb the inputs to see how the output changes.
Hi Mostafa Kiani Shahvandi ,
Thank you for the reply.
Yes, it's an input to an RNN network. Can you elaborate on perturbing the input, how can I achieve it?
For this kind of time series, there would be no ordinary type of feature selection. Two problems can be discussed. (1) In order to see what features are more important, you can add some sort of noise or outlier to data, then see what feature, when disturbed, changes the output more. That would be the most important feature and you can select it. (2) Feature extraction in the meaning of generation of new features can't be simply done.
Mostafa Kiani Shahvandi Referring to (1), for categorical features, does perturb mean including or excluding the whole one-hot encoded categorical feature to see the impact of the feature on the prediction output?
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python