some questions about RM

dan76

hi there,

i have some question and i think one of you can answer it surely.

1. the SOM-plot: could you tell me the decrease rate of the learning factor and the radius factor or can i read it somewhere?

2. is there an operator, that shows the silhouette-coeffizient and/or the compactness of the clustering eg kmedoid?

thanks for help.

dan

IngoRM

Hi,

1. the SOM-plot: could you tell me the decrease rate of the learning factor and the radius factor or can i read it somewhere?

I had to make a look into the code - which would also has been possible for you 

However, the adaptation is done via

learnRateCurrent = learnRateStart * Math.pow((learnRateEnd / learnRateStart), (((double) time) / ((double) maxtime)));

with learnRateStart = 0.8 and learnRateEnd = 0.01.

The radius is adapted by

adaptationRadiusStart * Math.pow((adaptationRadiusEnd / adaptationRadiusStart), (((double) time) / ((double) maxtime)));

with start = 5 and end = 1.

2. is there an operator, that shows the silhouette-coeffizient and/or the compactness of the clustering eg kmedoid?

I do not get the point here. If the quesiton is: "is there a kmedoid?" the answer is "yes". If the question targets at cluster evaluation methods, you will find those in the "performance" operator group.

Cheers,
Ingo

dan76

Thanks for the fast reply.

I run 2 kmedoids clustering one with k=5 and one with k=8.

I dont know as i can compare the quality? I have read somewhere that the quality of a cluster is defind as the sum of the distance of all members to the medoid. And the quality of the clustering bei the sum of the clusterquality. That is im looking for.

I hope the description fit

thanks again.

Dan

IngoRM

Hi Dan,

as I said, you will find operators for calculating the performance of clustering (including the "inner cluster density" like you described it) in the group "Validation" - "Performance" - "Clustering". Here you can use for example the operators "ClusterCentroidEvaluator" or "ClusterDensityEvaluator".

Cheers,
Ingo

dan76

Hi Ingo,

thanks, thats it.

greetz

Dan