2
$\begingroup$

I have a dataset collected using an accelerometer. I am extracting the magnitudes from the signal to find the difference in running pattern between two different running surfaces. Will normalization help to improve my classification accuracy? In general, when should we normalize time series data? Any suggestions?

$\endgroup$
  • 1
    $\begingroup$ I don't think normalizing the actual signal will help, although some classifiers have problem with different scales on the dimensions for the feature space. What are you using as features? $\endgroup$ – SlimJim Jan 25 '14 at 3:31
  • 1
    $\begingroup$ A common thing to do for the features space is normalize->PCA->train $\endgroup$ – SlimJim Jan 25 '14 at 3:32
  • 1
    $\begingroup$ Okay. The positive and negative magnitudes with respect to the x y and z axis and speed of the runner. Totally my feature set consists of 7 features. $\endgroup$ – user3088897 Jan 25 '14 at 3:36
  • $\begingroup$ I've seen 'normalize' used to mean too many different things (to standardize components, to scale to constant vector length, to scale components to [0,1], and even to transform to normality) to be confident I understand the question. I expect you mean vector normalization in this case (Euclidean norm?), but can you make that term more precise so that it's not just an assumption? $\endgroup$ – Glen_b -Reinstate Monica Feb 3 '14 at 21:38
3
$\begingroup$

Data normalization (centering & scaling) tends to helps more with model convergence/stability when dealing with maching learning algorithms. Feeding ML algorithms input data with wildly different mean/variance can slow or prevent model convergence.

Normalizing your data is also helpful in that it will make your model results (e.g.-regression coefficients) more easily interpretable.

I wouldn't expect that normalization would significantly improve your actual model results, but it couldn't hurt!

$\endgroup$
  • 2
    $\begingroup$ 1. Most (all?) ML algorithms which do not explicitly require center/scaling as a preparation step will not be alleviated of any convergence issues by centering/scaling. 2. Normalized data leads to less interpretable model coefficients, since e.g. Pearson correlation is a unitless quantity and hence has no contextual importance. I suspect only PCA and related variance decomposition algorithms require some form of normalization. $\endgroup$ – AdamO Feb 3 '14 at 21:03
  • 2
    $\begingroup$ The Gradient Descent Method is one such example where center/scaling is not explicitly required, but unbalanced input variances can lead to convergence problems. (The algorithm can "overshoot" the global minimum point of the optimization function.) As for interpretability, I suppose that's a bit of a give & take. What I had in mind was that with normalized input data, one can more easily compare which variables have the greatest impact on the Dependent Variable. $\endgroup$ – Jacob Feb 3 '14 at 21:10

Your Answer

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