# Clustering related areas with k-means

I am trying to cluster related areas of knowledge based in publications.For example a researcher has 3 keywords in a paper, in another paper he has 5 keywords, but 3 keywords are the same in the both papers. Then these 3 keywords are similar and It could be an area of knowledge. A sample of my file is:

'Domain Ontology,Semantic Web'
'Linked Data,Domain Ontology,Use Case'
'Domain Ontology,Linked Data,Semantic Annotation'
'GIS,Open GIS,Integrated Geo Systems'
'Open GIS,GIS'


My file contains 48963 rows and 19000 keywords. I have tried grouping words (like in the sample), using StringToWordVector (STWV), but I don't have good results. So I tried different K for my cluster since 3 until 13000. When K is greater my log likelihood decrease. In many models my words are grouped perfectly, but the percentages are not really good for some clusters. For example for K=3000, the 2999 have between 0,1% and 1% of the data, but the 3000 have the 21%. I am using WEKA. Anyone know some work related, some advice or any help is helpful...

Cheers.

• These results sound typical for me, with k-means on text. Text is not what k-means was designed for, and cosine is okay for text search, but it doesn't quantiy similarity in a reliable way. – Has QUIT--Anony-Mousse Oct 11 '15 at 20:24
• @Anony-Mousse So then Is there another reliable way to build my cluster? Or How do I estimate my true K? – c.uent Oct 12 '15 at 2:11
• There probably is no true k. And don't expect clustering to be reliable: it is an explorative approach, not an automatic. – Has QUIT--Anony-Mousse Oct 12 '15 at 5:43
• Yes your right! But I'm using a bag of words, so at least I could cluster similar words based in the concurrency of different papers. For example: Domain Ontology GIS Integrated Geo Systems Linked Data Open GIS Semantic Annotation Semantic Web Use Case Cluster 1 0 0 1 0 0 1 0 cluster1 1 0 0 1 0 0 0 1 cluster1 1 0 0 1 0 1 0 0 cluster1 0 1 1 0 1 0 0 0 cluster0 0 1 0 0 1 0 0 0 cluster0 Am I wrong? – c.uent Oct 12 '15 at 22:29
• Good luck. Language when you only see bits and bytes is far from easy. I doubt 10% of your cluster assignments will be useable. – Has QUIT--Anony-Mousse Oct 12 '15 at 22:47