What threshold value that you need to fix? Please elaborate on that.

HTH.

Samer

If i select canopy clustering with k-means clustering in weka , it is necessary to give T1 & T2 values as a threshold values, i am not getting how to calculate and give that value.

For T1, you can either supply a value > 0, and like T2 it is used as-is, or the absolute value of a value < 0 is used as a multiplier of T2 to set the value for T1.

There are no real guidelines on how to set T2 and T1. You should just experiment to see how different values affect the result. One thing you could try is to wrap Canopy in the MakeDensityBasedClusterer. Then you can take a look at the log likelihood of some test data (using a percentage split or separate test set) in order to see the effects of varying the values of T2 and T1.

Cheers, 

Samer 

The most effective way is experimental analysis. But for prove it experimentally, you should plot ROC curve. ROC curve scan show you how the changes in the thresholds affect you final results. Then, you cut-off values for threshold would be the points that you wouldn't have significant changes in performance metrics (e.g. accuracy and false alarm rate). For instance, accuracy is reasonably high and false alarm rate is reasonably low.

Canopy is a clusterer algorithm, (i.e., unsupervised learner). ROC curve, accuracy, etc. can't be applied, though. In order to evaluate Canopy, we rely on the "log likelihood" aspect. 

Cheers, 

Samer

You are right, but that is not true in all cases. ROC curve can be applied for optimizing any threshold. We do not know about the details of study, so for excluding ROC as a choice, we need more details.  Besides, sometimes accuracy can be applied to clustering. Although the main advantage of unsupervised learners is working with unlabeled data, sometimes we have labeled data but we do not want to use it in data analysis. Since it the real application we do not have labeled data, but for research purposes we have some labeled data. So we assume unlabeled data and use labeled data only for performance evaluation. Intrusion detection, the KDD cup dataset, probably the dataset that she is using is one of those application. We have labeled data for research purposes.     

Generally, ROC curve requires labeled data and "Canopy" is a clusterer method--deals with data that have no labels. 

In WEKA ( Vinutha uses), the ROC curve and another measures such accuracy, F-measure, etc. can't be applied with a clustering method whatever her dataset structure; therefore, we, in WEKA environment, don't evaluate the cluster technique using, e.g., ROC curve, unless this technique gets converted to a classification process.

Cheers,

Samer

Thank you Samer and Vahid.

You are right Samer if we use ROC curve it gets converted into classification.