Search code examples
javahierarchical-clusteringelki

The use of ELKI source code for hierarchical clustering

I try to use ELKI (Environment for Developing KDD-Applications Supported by Index-Structures) for hierarchical clustering. So some days ago, I imported the ELKI source code (Maven projects) and then I run the miniGUI.

My questions are:

  1. I want to use ResultVisulizer for result. However, I found that there is no ResultVisulizer item in the resulthandler. What should I do to make this ResultVisulizer item to be listed here?

  2. I want to see the dendrogram of hierarchical clustering. Which parameters need to be set in this miniGUI for this purpose?

Solution

    1. To enable the SVG visualizer, enable the svg maven profile. SVG adds about 17 MB of dependencies, so we decided to make this an optional module.

    2. There currently is no dendrogram visualization in ELKI, because such visualizations are not very scalable; with more than a few dozen instances they become hard to read. Also, on many interesting data sets the distances (heights) are too similar, so that the dendrogram again is not very useful, unfortunately.

      If you extract clusters and simplify the hierarchy, ELKI will display the cluster hierarchy tree, but it will not use the height information (for the reasons above, and for readability reasons). ELKI has powerful cluster extraction methods from hierarchical clustering that are not available in R.

      You are welcome to contribute a dendrogram visualization; but please try to make it useful for a wide user base without the need of too much customization.

    Here is an example from R (Source) of why I'm not convinced of dendrograms (and why it is low on my personal priority list - someone else will have to contribute this to ELKI): Dendrogram fail