Shifting and scaling patterns from gene expression data.

Area of Science:

• Data Mining
• Bioinformatics
• Machine Learning

Background:

• Biclustering is increasingly popular for high-dimensional data, extracting clusters using subsets of attributes.
• It handles overlapping patterns, offering advantages over traditional clustering.
• The mean squared residue is a common measure for bicluster discovery since Cheng and Church.

Purpose of the Study:

• To evaluate the effectiveness of the mean squared residue for identifying shifting and scaling patterns in biclustering.
• To mathematically prove the limitations of the mean squared residue for simultaneous discovery of both pattern types.

Main Methods:

• Analysis of biclustering algorithms utilizing the mean squared residue.
• Mathematical assessment of the mean squared residue's performance with shifting and scaling patterns.
• Investigation of the measure's sensitivity to gene value variance.

Main Results:

• The mean squared residue is effective for shifting patterns but not for scaling patterns.
• A perfect scaling pattern does not yield a zero mean squared residue.
• The measure's dependency on scaling factor variance can hinder the detection of scaling patterns, especially with high gene value variance.

Conclusions:

• The mean squared residue is mathematically imprecise for simultaneously discovering shifting and scaling patterns.
• Future biclustering algorithms may require alternative measures to accurately detect both pattern types.