Biclustering: methods, software and application

Más información sobre el libro

Over the past decade, biclustering has gained traction in biological data analysis and other areas involving high-dimensional two-way datasets. This technique simultaneously clusters both rows and columns, identifying subgroups of objects that share similarities in specific variables while differing in others. This dissertation aims to enhance biclustering methods, addressing the sensitivity of existing techniques to parameter variations and data fluctuations. An ensemble method was developed to improve stability and reliability, allowing for more consistent bicluster retrieval through varied parameter settings or by utilizing sub- or bootstrap samples of data. A software package was created, featuring a collection of bicluster algorithms tailored for diverse clustering tasks and data scales, along with new visualization methods for bicluster solutions. Traditional cluster validation indices, such as the Jaccard index, were adapted for the bicluster framework. The research also applied biclustering to marketing data, adjusting established algorithms for different data contexts and developing a new method for ordinal data. To validate this approach, artificial data with correlated random values was generated based on a probability vector and correlation structure. All methods discussed are accessible in the R packages biclust and orddata, with numerous examples provided to demonstrate their application.