1
$\begingroup$

I'm having a hard time getting kmeans to cluster data effectively. It fails to segment data well even for a simple attribute with 5 categories. I'm aware of DBSCAN, Hierarchical Clustering and GMM. However, just wanted to know if there's any way (visual or otherwise) to narrow down the clustering algorithm which might work on the dataset in question, before I start to write the code for each of these algorithms.

Thanks in advance.

$\endgroup$
1
$\begingroup$

No.

Clustering is an explorative technique, it is subjective what is good, and the best clusters are those that are "interpretable but unexpected", a property that you cannot quantity with statistics. So it is a trial-and-error task.

Furthermore, data preparation is much more important than the choice of clustering algorithm. On badly prepared data, none will work.

Last but not least, categoricial data is a huge problem. It lacks detail for most clustering approaches - treating this as binary variables is much too coarse and tends to produce bad solutions (such as tiny "clusters" and trivial splits on a single variable). This is likely a problem of the data, not the algorithm. Similar issues can be seen with integer attributes or any other attribute that has only very few discrete levels (including Likert-like-scale questionnaires). Methods such as k-modes exist for categoricial data, but often don't produce better results either...

| improve this answer | |
$\endgroup$

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.