Similarity Measures for Comparing Biclusterings

Area of Science:

• Data Mining
• Machine Learning
• Bioinformatics

Background:

• Clustering partition comparison is well-established, with numerous studies on similarity measures.
• Existing similarity measures for clusterings are not directly applicable to biclusterings due to their structure (tuples of row and column sets).
• Current biclustering similarity measures are often minor contributions, overlooking desirable properties.

Purpose of the Study:

• To review existing biclustering similarity measures.
• To define and discuss eight desirable properties for biclustering comparison.
• To evaluate which reviewed measures possess these properties and demonstrate potential misleading evaluations.

Main Methods:

• Comprehensive literature review of 14 biclustering similarity measures.
• Formal definition and theoretical analysis of eight desirable properties for biclustering measures.
• Empirical analysis with examples to illustrate the impact of missing properties on evaluation accuracy.

Main Results:

• Identification of significant gaps in the desirable properties of existing biclustering similarity measures.
• Demonstration of how specific measures can yield misleading results in important biclustering studies.
• Validation of the importance of the defined properties through illustrative examples.

Conclusions:

• Existing biclustering similarity measures often lack crucial properties, leading to unreliable evaluations.
• A more general comparison approach, transforming biclustering comparison into overlapping cluster comparison, is advocated.
• Further research is needed to develop and validate biclustering measures with the identified desirable properties.