** Given:
a) Ground truth clusters for a data,
b) Clusters obtained using a clustering algorithm (eg: DBSCAN) when applied on the data after processing it .
** Issue:
How to evaluate the performance of the clustering technique when applied on a specific data??
** NMI (Normalized Mutual Information) is a popular external measure to do so. But in cases like below, it gives bad results:
E.g:
Ground_truth = [1,1,1,1,1] ;
DBSCAN_Clusters = [1,1,1,1,2];
nmi = normalized_mutual_info_score(Ground_truth, DBSCAN_Clusters); %python code
** The value of the variable "nmi" approximately equal to zero in this case.
** Here, note that, nmi = 0 in-spite of the fact that DBSCAN (clustering algorithm) has failed to cluster only one cluster member and rest four matches the ground truth.
** This is a typical case when the ground truth contains only one cluster.
** Questions :
1) Why does this happen?
2) Does it mean that clustering algorithm is performing bad?
3) Should I use other measures along with NMI ? If so which ones, and what are they for?
Thanks.
It is but it depends on what is your point of improvement from the baseline paper or even if it is entire a novel proposed method.
Desmond Bala Bisandu : Thanks for the answer.
But, I was trying to ask a general question.
Let me put it in a simple way:
"Why does NMI value becomes close to zero when the ground truth has only one cluster + when my clustering result is not exactly matching the ground truth?
In such cases what are the other options to evaluate my clustering technique? "
I think the following site could be useful for you:
https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-clustering-1.html
hoping it answers your question
I think you can add other measure to be able to discuss the obtained result such as the error measure, the F-measure and if you are working with density based clustering you can add internal evaluation metrics such as cd_bw index and DBCV index.
hoping it helps you in your question.
Thanks @Yusra and @Soumaya ..
I hope I can find what I am looking for combining both your answers.
Read this https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-clustering-1.html
- Badges
- Science topic