Center for Statistical Sciences Seminar
Abstract:
Since the advancement of the microarray technology, various statistical
techniques have been applied or developed for analyzing the data. Unsupervised
analysis (e.g., clustering) plays an important role among them. For instance,
clustering algorithm is commonly used in the discovery of cancer signatures by
revealing "hidden" clusters that might be relevant to tumor risk structures.
However, like many other statistical techniques, these results are usually
dependent on the techniques employed and difficult to replicate with different
clustering methods.
Some recent researches are trying to deal with this issue by perturbing the
microarray data randomly and then assessing the reliability with perturbed data.
Commonly used techniques include bootstrapping, noise injections, random
subspace and random projection. In this study, we focused on the random
projection technique due to its superiority in the theoretical supporting and
noise injection controlling. We first develop a method to set up the cutoff
that can be used in the random projection to distinguish stable and unstable
clusters (or samples). Then we establish an approach to select the optimum
structures after data perturbation. We also evaluate the relationship between
clustering reliability and some QA/QC metrics in microarray experiments. Now
we are studying the effect of various agglomeration methods on the reliability
assessment in microarray clustering.
PDE Seminar
<--- 2008 Index