Seed Initialization Method for Hierarchical Clustering

[octavianidevi]

Dear all,

I am a bit confuse about how cluster node in sas miner handle k-means and hierarchical clustering.

I read book 'Data Mining using SAS Enterprise Miner', it says that "The number of Cluster option actually determines if you would like to perform either hierarchical or partitive clustering(K-means). Hierarchical clustering can be performed by selecting the Automatic option.... Conversly, selecting the User Specify option perfors partitive clustering ". When we choose Number of clusters 'Automatic', we can choose clustering method under selection criterion either ward, average or centroid.

My confusion here is, if I choose Number of clusters 'Automatic' which mean that I perform Hierarchical Clustering then what is the use of Seed Initialization Method ? I have tried that different Seed Initialization Method gave different number of clusters on automatic mode. It is confusing since as I understand, hierarchical clustering doesn't initialization seed.

Anybody can explain about it? Thank you

[mohamed_zaki · Accepted solution]

First if this book is Randall Matignon book then this book is old and based on Miner v4. It still may be helpful to understand more about Miner, but i am not sure if it is the best resource to learn.

Tip: Guidelines for Choosing a Clustering Method in the Cluster Node ,  this articale give good explanation of the automatic selection process. 

It is not k-means and hierarchical clustering. What happen is based on the number of seeds (50 by default) training data are distributed to the closest seed. Then the means of these intial clusters are calculated. After that the hierarchical clustering consolidate these clusters and within the CCC is calculated. Finally the final number of cluster provide the K in K-means and the clusters are obtained using a k-means algorithm.

So yes, the number of seeds affect the final number of cluster in the automatic selection process.

[VanDalucas]

Hi Mohamed, correct me if I am wrong

Let's say the ward algorithm is completed and using the CCC criterion we find out that K=5. Since ward method acts by merging clusters (starting by assigning one cluster to each observation) then at some point there must have been k-clusters. After K-1 repetitions of merging we are left we 1 cluster, at this point the ward algorithm has done its job and K-mean is initiated.

My question is: Does the K-mean algorithm to create the final number of K-clusters takes into accout the K clusters that were created K-1 repetitions before the end of the ward method or does is randomly chooses K observations as seeds for the final K clusters

Thank you