I have a set of real data containing labels. However, in some biclustering algorithms (e.g., Cheng and Church Algorithm), originally authors applied gene expression data like Yeast without having labels. My question is that what are appropriate evaluation measures in this scenario?
Briefly evaluation measures, which I presume means "features", are determined using supervised learning using a technique called "feature selection". Unsupervised learning generally is used to identify categories for a supervised learning algorithm. The orthodoxy has changed.
"Unsupervised learning feature selection"
- To identify suitable features using unsupervised learning a priori then genome location which must have a major outcome, thereafter it would be a case of deleting a given feature and re-evaluating the clusters. If the clusters disintegrate then that is a key feature (not very cool but would work for datasets with limited numbers of features).
- Weighted unsupervised learning is popular, this is basically "feature selection", i.e. which features contribute to the clustering. This is basically the formal version of point 1. Personally I have had little success with this approach, i.e. the results were not informative, but it does work. The results are not discriminatory between different conditions, which is what I was seeking.
Regarding labels, labels are not used in unsupervised learning approach - except weighted unsupervised learning - hence their presence or absence usually makes little difference. The only "label" used is the sample type, i.e. the "id" and not the "feature". Without intrinsically understanding the experiment it is difficult to forward an accurate a priori assessment.