1
$\begingroup$

I am having a project in which I need to group test cases failing due to same faults, and obviously, test cases are not labeled with due-to fault. So clearly we have an unsupervised classification (clustering) problem. However, the requirement is that we need to classify new coming failed test cases into existing/new groups.

My question is: can we use the result of cluster analysis (e.g. K-means) as the input to train a classifier (e.g. Naive-Bayes or SVM) for future prediction; we also plan to implement a feedback system (let the users tell whether the prediction is right or wrong and let them suggest the correct prediction) for classifier re-training.

Or it is better to use only the resulting clusters to predict new data (decide which centroid new data belongs to, use HDBSCAN for example)?

I appreciate all your answers and suggestions.

Thanks in advance.

$\endgroup$
1
  • $\begingroup$ It's just speculation that the clusters will have any relationship with future faults. I doubt this approach will work. Instead, get some labeled data first, then train a classifier; or use one-time lass classification to detect change. $\endgroup$ Mar 24, 2019 at 16:47

1 Answer 1

-1
$\begingroup$

I think it's okay to use the output of a clustering algorithm as an input to a classifier. Reason being, k-means is a partitional clustering algorithm. It will help you group your data. But remember, k-means works only for numerical data. So if your data got both numerical and text values, then a common method is to discretize the text/categorical values to numeric and then apply k-means.

Moreover, if clustering is a pre-processing algorithm. So, if you are data is unlabeled, you can use clustering to find patterns or groups. Thereafter, you can manually label the group's basis of the type of values contained in the group. And then apply a classifier to do prediction.

$\endgroup$

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.