Hi Jose, you should have a look at feature weighting clustering approaches. A very good review is given in

de Amorim, R. C. (2016). A survey on feature weighting based k-means algorithms. Journal of Classification, 33, 210-242.

The R package 'wskm' implements the following methods

- Liping Jing, Michael K. Ng and Joshua Zhexue Huang (2007). An Entropy Weighting k-Means Algorithm for Subspace Clustering of High-Dimensional Sparse Data. IEEE Transactions on Knowledge and Data Engineering, 19(8), 1026–1041

- Xiaojun Chen, Yunming Ye, Xiaofei Xu and Joshua Zhexue Huang (2012). A Feature Group Weighting Method for Subspace Clustering of High-Dimensional Data. Pattern Recognition, 45(1), 434–446

- Xiaojun Chen, Xiaofei Xu, Joshua Zhexue Huang and Yunming Ye (2013). TW-k-Means: Automated Two-level Variable Weighting Clustering Algorithm for Multiview Data. IEEE Transactions on Knowledge and Data Engineering, 25(4), 932–944.

but these are partitional algorithms. I also implemented a weighted version of

Park, H. S., & Jun, C. H. (2009). A simple and fast algorithm for K-medoids clustering. Expert systems with applications, 36(2), 3336-3341.

but it does not provide cluster-specific weights for each variable, so the algorithms included in the 'wskm' package probably fit better with your question.

Dear Raffaele Mattera,

I had a look at the package you recommended, and I have another question about it (the manual is not very explicative, and I couldn't find the vignette of the package either).

Is it possible to use a pre-computed hierarchical cluster (hc) object as input to the module`s functions? I say this because I have already done the rest of the calculations on my hc object, and I do not want to recalculate it in another function since the branches may vary and ruin the rest of the analysis...

Thanks in advance & best wishes

JL

Hi Jose, I don't think so. Weights are optimally computed within the algorithm with the aim of minimizing the objective function in an iterative manner. Said differently, the weights are used to define the partition.

You can still perform some analyses at the cluster level and try to understand what are the features "better representing" the units in each group. In this case, however, any feature relevance argument is ex-post and it is not considered in the clustering process (during clustering you implicitly consider all features to be equally important instead).

Raffaele Mattera , what do you think about this?

It looks cool, but still does not do exactly what I was looking for.

https://rpubs.com/pjmurphy/269609

the higher the variability of variable values by class (category), the more valuable it is to separate them