0
$\begingroup$

Suppose I have a dataset containing two very similar classes of data. By similar, I mean that the 'distance' between these two classes is very small. For example, one instance in Class 1 is the sum of one instance in Class 2 and some noise. If the SNR is high, we can say these two instances are similar. Because of the similarity, the dataset is inseparable. I am wondering if there exists any effective clustering algorithm that can work. Thank you very much!

$\endgroup$
3
  • $\begingroup$ Welcome to DataScienceSE. Logically if instance A is the same as instance B + noise then they should be considered identical, therefore in the same cluster. If all your cases are like this then clustering is not what you need. Otherwise it's matter of defining the right custom distance. $\endgroup$
    – Erwan
    Feb 12, 2021 at 12:06
  • $\begingroup$ @Erwan Thank you for your reply. I give this example only to show the inseparability. Based on your remark, you mean that no clustering can work for the inseparable data, don't you? $\endgroup$ Feb 13, 2021 at 2:02
  • $\begingroup$ It depends what is the goal, but clustering is about grouping and "separating" instances so it's difficult to imagine a case where it makes sense to separate things which are inseparable. Maybe if you could give more detail about the context it would help. $\endgroup$
    – Erwan
    Feb 13, 2021 at 11:09

1 Answer 1

3
$\begingroup$

As you point out, the problem is not on the clustering algorithm, but on the features. So the question comes to the particular data you might be dealing with.

As an example, say you want to cluster different kind of animals. It is in general much easier to tell an elephant from a horse apart. But if you want to distinguish between races of horses, it gets much harder. But the bottleneck lies (mostly) on the features.

$\endgroup$
2
  • $\begingroup$ Thank you for your reply. You mean I need to do feature extraction to make the new data separable, don't you? $\endgroup$ Feb 13, 2021 at 2:34
  • $\begingroup$ You need to extract good features that help you separate clusters better. Of course, the loss function also plays an important role, but features carry the heaviest load for this task. $\endgroup$
    – jpmuc
    Feb 13, 2021 at 9:06

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.