3
$\begingroup$

I want to do some clustering for a dataset where I am looking at 10,000 peoples usage of certain electronic devices. I have 11 columns; the first column is simply a URN representing each person in the study. Then the other ten columns are for the weekly usage, in minutes, of each device (e.g. phone, laptop, t.v. etc).

No individual actually has all ten devices. They normally tend to have 2-4. For this reason they will have a value of n for each of the devices they use, and then a "0" if they do not have said device. Therefore, I obviously can't omit rows which have a 0 in them, and the 0 does not necessarily mean missing data, as opposed to actually meaning they do not own the device.

I am wondering how a typical basic cluster analysis would be carried out in a situation like this? / what would be an appropriate way to approach such a study? (I am using R for the analysis, for reference)

Thanks for the help.

$\endgroup$
2
  • 1
    $\begingroup$ You can just convert the columns to Categorical and cluster. Or Replace 0 with any other arbitrary number and convert the column to categorical before applying the clustering. $\endgroup$
    – Syenix
    Dec 17 '19 at 6:08
  • 1
    $\begingroup$ @Syenix, thank you. I am definitely going to be doing an approach where I make the data binary to represent "does/does not" have the device, so along the lines of what you have said. But surely there is value in using an approach whereby the variables are numeric and i am keeping in the "minutes used" for each device? In this case, I am still not sure if i should leave all the zeros, or make them '1' , for example, and how this will affect my results. $\endgroup$
    – calmac07
    Dec 17 '19 at 16:48
1
$\begingroup$

K-means is the wrong algorithm for this problem.

It assumes dense, continuous input data, which you haven't.

$\endgroup$
2
  • 1
    $\begingroup$ Would you know what would be an appropriate alternative? Thanks $\endgroup$
    – calmac07
    Dec 18 '19 at 0:50
  • 1
    $\begingroup$ Hierarchical clustering for example. Maybe a divergence measure is a suitable distance. $\endgroup$ Dec 18 '19 at 6:29

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.