Biclustering with missing data

Abstract

Biclustering is a statistical learning methodology that simultaneously partitions rows and columns of a rectangular data array into homogeneous subsets. Biclustering is known to be an NP-hard problem, and therefore various heuristic approaches have been proposed. These strategies break down when dealing with any degree of missing data in a two-way table of data values. To address this issue, we propose a new biclustering method based on the work of Li in 2014 [18]. Numerical results show our approach performs well on moderate-sized test cases with even a large missing-value percentage (99%+). To illustrate the practical usefulness of the method we provide two case studies. The first is an agricultural application where rows represent plant varieties, columns represent planting locations, and data are yield values. The other is a well-known movie rater application where rows represent raters, columns represent movies, and data are ratings.