In many instant it has been said cutting a dendrogram at a certain level gives a set of clusters. Cutting at another level gives another set of clusters. How would you pick where to cut the dendrogram?
Is there something we could consider an optimal point? I have also wondered about this problem, but (unfortunately) haven't found any convincing answers yet.
So is that correct to say “there is no definitive answer since cluster analysis is essentially an exploratory approach; the interpretation of the resulting hierarchical structure is context-dependent and often several solutions are equally good from a theoretical point of view”. Please help me.
Well, I guess it is depended the natural of the data set. I assume you are using hierarchical cluster analysis? The dendrogram is very flexible to be explained, which requires the domain knowledge of the data.
If you want more solid evidence about whether a cluster number is good or bad, go for silhouette analysis or BIC score.
If you don't know the initial number for clustering, I suggest you go for Variational Bayesian Gaussian Mixture. It will output a cluster number that is proper based on the input data.
Personally, I suggest you go for Latent Tree Model clustering, this type of clustering analysis will give you a tree structure with a set of reasonable and possible clustering strategy (the number of clusters and the weight between them).
Thank you Hassen Beshir and Ten Fu for your fast response.
@ Teng Fu: Thanks again, now I have ideas on the advantages of Latent Class cluster models over more traditional ad-hoc types of cluster analysis methods include model selection criteria and probability-based classification. Furthermore, it is possible to include variables of different scales (continuous, ordinal or nominal) within the same model.
But this can be done on XLSTAT software, which I couldn’t get. Can Latent Class cluster be work out in STATA, SPSS,SAS? or any suggestions?
@ Teng Fu: Good information, unfortunately, my data is continuous (It is efficiency (E) scores, ranging 0
Hi, you can use very simple "Elbow rule", see: https://en.wikipedia.org/wiki/Determining_the_number_of_clusters_in_a_data_set
In brief, when you have k clusters and adding another k+1 cluster does not increase the fit of the model significantly, you stop adding clusters.
Hi Beneberu
It's not an intuitive issue since cluster is unsupervised learning technique. However, you may apply Expectation–Maximization (EM) algorithm that internally uses cross-validation to determine cluster number. On the other hand, if your data have class attribute, then you could rely on this class to cluster the data.
HTH.
Samer
- Badges
- Science topic
- Similar topics
- Biological Science
- Microbiology
- Microorganisms