I have a set of data and I want to know that whether they are necessary to add in the clustering analysis. Like ONEOFF_PURCHASES_FREQUENCY, I am not sure it is wether helpful in doing cluster analysis.



1 Answer 1


Clustering may be useful if your dataset consists of distinct groups where each group behaves differently. For example, you have different types of customers, and each type tends to prefer different products and shows different buying patterns. In that case, after splitting the dataset into clusters, subsequent classification or regression tasks can be performed differently on each cluster, so that trends unique to each cluster can be learned.

The obvious way to see, if clustering is useful is as follows. Cluster your dataset choosing the number of clusters either from the domain knowledge or by clustering algorithms (see e.g. this, this, and this), then do your classification or regression task on each cluster separately. Evaluate it with some metric, e.g. $R^2$, $RMSE$, accuracy, $F_1$-score, etc. Then calculate the same metric for the whole dataset without clustering, and see if clustering improves your metric significantly.

It would be great, of course, to know how to estimate the usefulness of clustering before doing classification/regression task but I don't know such methods. What comes to mind first is to do exploratory data analysis for each cluster separately and see how it differs from the same analysis for the whole dataset. However, the straightforward method above might be faster, especially if you use some simple algorithm (e.g. decision tree) just for that estimation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.