For training a Hidden Markov Model (HMM) on a multivariate, continuous time series, is it preferable to scale the data somehow? Some pre-processing steps may be:

  1. Normalize to 0-mean and unit-variance
  2. Scale to [-1, 1] interval
  3. Scale to [0, 1] interval

With neural networks, the rationale behind scaling is to get an "un-squished" error surface that is easier to navigate in.

HMMs use the Baum-Welch algorithm, which is a variation on the Expectation Maximization (EM) algorithm, to learn parameters.

Is EM sensitive to scale of features? Is there some motivation for normalization for HMMs?



Your Answer

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