I have a dataset that's purely categorical: for each item it's ranked across a set of attributes, whether it's easy, moderate or difficult.

But there are blanks if the item doesn't have the attribute. Currently I intend to encode blank as 0 if that's appropriate, and difficulty level as 1, 2, 3.

In this case, should I use Frequent Itemset Mining (Apriory) or k-modes if want to pinpoint which items are the most similar & dissimilar, and how could I identity what kind of attribute combinations result in such clustering? and how do I assess if the clustering result is good or need further tuning?

  • $\begingroup$ Notice that prior to encoding missing values with 0, your easy-moderate-difficult is a typical 3-point Likert scale. Thus, you have much more analytical power, even possibly assuming that it is interval, which is a very common assumption with Likert-scale data. An alternative option you have for missing data is to consider the frequency of the gaps and see if you can work with a subset of the data (be it a restriction on data points or attributes) that is more complete. $\endgroup$ – mapto Aug 9 '18 at 15:22
  • $\begingroup$ Another technique to deal with the missing data is imputation. I'd argue that it is still better than loosing your total order when encoding blanks as 0s. $\endgroup$ – mapto Aug 9 '18 at 15:24
  • $\begingroup$ As for directly answering your question, I suspect the answer would be "it depends" and "not enough information". It is rather a case of trying both and seeing how they perform. This shouldn't be too difficult to do with any standard package that you might be using. $\endgroup$ – mapto Aug 9 '18 at 15:26

Your Answer

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

Browse other questions tagged or ask your own question.