I'm currently perfoming an experiment with a component of human based computation. People should try to cluster images generated from gene expression (Self Organizing Maps) using their own perceptual capabilities I'm gathering data in a fashion intended to resemble a hierarchical clustering, At the end, every person should classify thirty images in six groups of five members according to the perceived similarity (perhaps having a better performance than usual algorithms due to Gestalt perception). The thing is that at the end I'll have many sets of images arranged in a particular way for each person (but always 6 groups and 5 members), and I'd like to have a consensus cluster to contrast with computational methods and with the real known labels of the data. I'd appreciate any suggestion in how should I generate this consensus clusters, namely algorithms, programs, papers.
Thank you beforehand and do not hesitate to ask for additional details if needed.
Best regards.
Daniel.
So the constraint here is that you must have 6 groups of 5 members (I take it there is good reason to believe that).
My first inclination is to explore the space of candidate consensus group assignments and minimise the sum of edit distances of each subject's group assignments to the nearest group. A sort of modification of the ideas behind nearest centroid clustering. I think a reasonable solution could be obtained by a walk through the assignment space pretty quickly. The code to do so could be knocked together in about an hour in R. The final assessment would be edit distance between your consensus and the ground truth. (You can use some other metric than edit distance too).
I wouldn't do any form of hierarchical clustering - you'll never get the the required structure that you insist on.
Hi, David. Thanks for your answer. I've been actually reconsidering the real necessity of the structure of 6 groups and 5 members. Initially I constrained it to 30 randomized SOMs thinking in a reasonable size for human analysis, and I assured to choose 5 members of the 6 molecular subtypes that I know beforehand, so I thought that steering it to 6 groups and 5 members each could help, but now I think it has no real benefit, even considering that people struggle to achieve that particular structure.
I've been considering using Cluster-based Similarity Partitioning Algorithm (CSPA) from Strehl and Ghosh http://strehl.com/download/strehl-jmlr02.pdf
This algorithm has very basic heuristis, consisting in, basically, reverse engineer the individual cluster, binarizing the data, giving 1 to those elements that belong to the same group in that cluster, and 0 if not. This creates a similarity matrix that later can be averaged for all the clusters and applying any other clustering method, achieve the consensus cluster.
http://strehl.com/diss/node80.html
What do you think about this, David?
- Badges
- Science topic