Iterated Matrix Reordering

Abstract

Heatmaps are a popular data visualization technique that allows to visualize a matrix or table in its entirety. An important step in the creation of insightful heatmaps is the determination of a good order for the rows and the columns, that is, the use of appropriate matrix reordering or seriation techniques. Unfortunately, by using artificial data with known patterns, it can be shown that existing matrix ordering techniques often fail to identify good orderings in data in the presence of noise. In this paper, we propose a novel technique that addresses this weakness. Its key idea is to make an underlying base matrix ordering technique more robust to noise by embedding it into an iterated loop with image processing techniques. Experiments show that this iterative technique improves the quality of the matrix ordering found by all base ordering methods evaluated, both for artificial and real-world data, while still offering high levels of computational performance as well.