RoBiC System Overview

The RoBiC system learns microarray classifiers by first reducing the dimentionality of data matrix using biclusters.
In general, a bicluster is a subset of genes and a subset of samples whose expression values have similar patterns; here, each bicluster is a sparse rank-one matrix -- ie, the outer product of two sparse vector.

TechReport (9page): "Using Rank-1 Biclusters to Classify Microarray Data"   (9/Apr/07)     

TechReport (7page): "Using Rank-One Biclusters to Classify Microarray Data"   (21/May/07)
    Nasimeh Asgarian, Russell Greiner

MSc Dissertation: "Using Rank-1 BiClusters", (Jan 2007; Nasimeh Asgarian)

Related Algorithms and Approaches

• Approaches related to RoBiC:
  * How RoBiC differs from SVD and related approaches
  * How RoBiC relates to other relevant results.
• RoBiC finds a set of BiClusters, then uses them to produce a classifier.
  This page presents other ways to use BiClusters to build a classifier

• RoBiC uses a particular "hinge" function to decide which patients and which genes belong to a bicluster.
  This page describes other "hinge functions"

• As suggested by the Figure above, we first form biclusters based on both test and (unlabeled) training instances.
  This page describes how this compares to simply adding a single test instance at a time (to the training set) when finding the biclusters.

Details of Empirical Studies

• Summary of the Data/Results on Prognostic datasets
• BiCluster Characteristics (Prognostics DBs)
  
  
• Summary of the Data/Results on Diagnostic datasets
• BiCluster Characteristics (Diagnostic DBs)
• Timing Information

Request

• If you have better results on these data sets, based on hold-out data (or CrossValidation), please email the relevant information
  
• dataset, approach, results
  to Nasimeh Asgarian.
• We are looking for other microarray datasets:
  • Binary classification labels (either included with each sample, or better: some withheld until we produce our predictions)
  • Complete data (ie, include a meaningful numeric score for each gene/sample pair)

  If you have such datasets, please send them to us. We will, of course, accept standard confidentiality agreements; just let us know.
  Our system uses data in the same format as the Plaid System; see here.
  (Of course, we can convert from other formats.)

Reading Group