0
$\begingroup$

I'm currently working with a large multivariate data set where I plan to use K-Means to try to find any associations in the data.

I'm not particularly well-versed when it comes to statistics, though I did realize I needed to exclude outliers from my dataset.

Assuming I have a 3 numeric variable dataset, would it be correct to:

  1. Just scale the data and remove outliers from there, and then K-Means.
  2. Or scale the data, remove outliers, then use the normal data set now excluded of the outliers and then K-Means.

Essentially the difference between the two is that in one I am working with a scaled dataset, and another I am working with just the normal data. Both are removed of outliers > 2 standard deviations out.

$\endgroup$
  • 3
    $\begingroup$ Why are you so sure you need to remove the outliers? You need to be careful following that line of thinking. Are the outliers incorrect data entry? Or are they data that violates your assumptions? If it's the second case, you need to think very carefully about just removing them. Can you post any examples of your data, indicating the outliers? $\endgroup$ – Conor Neilson Jul 24 '17 at 23:46
0
$\begingroup$

As @Conor Neilson mentioned, first we need to analyse the data and find whether the outliers are because of data entry error or not then check the proportion of outliers in the data set. If the proportion is low then we can remove those records from the analysis else replace them with central tendency measures based on the variable type. After taking care of outliers, we can to scale the variables.

In K-means there is no need to take care of outliers but scaling needs to be done as variables having large scales will inflate the distance between the data points.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.