I have a data set for second hand vehicle sales with 2 variables price and kms covered. If we plot these data in a 2-dimensional plane, the points will form scattered chart. I would like to select k samples (k - no of samples given by the user at run time) from the total data, maybe some 5,000 records. The samples selected must ensure that they are real representatives of the 5,000 records and the samples selected must be linked with the other underlying data. So when we select a particular sample, it should be possible to get the data for which the sample acts as representative. This may be done in more than one level also as done in subspace clustering.

Random sampling, density biased sampling or sorting by attributes will not help. Because a vehicle with more kms run may be priced high because it is in good condition. And a vehicle with less kms covered may be priced lower because there may some problem with the vehicle.

So, if there is any statistic method available to solve this problem please help me.

  • 1
    $\begingroup$ Exactly why won't random sampling work? It satisfies all your criteria. $\endgroup$
    – whuber
    Sep 19 '11 at 14:34

As whuber said, random sampling satistfies your criteria, so it should work.

Here are three alternatives that might be what you're looking for:

K-medoids and affinity propagation both guarantee that the examplars selected correspond to sample points from the data, whereas the means of clusters in k-means typically do not.


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.