Table cartogram generation as an optimization problem
Keywords: Table cartogram, Continuous area cartogram, Data visualization, Optimization
Abstract. A table cartogram, visualization of table-form data, is a rectangle-shaped table in which each cell is transformed to express the magnitude of positive weight by its area while maintaining the adjacency relationship of cells in the original table. Winter (2011) applies an area cartogram generation method of Gastner and Newman (2004) for their generation, and Evans et al. (2018) proposes a new geometric procedure. The rows and columns on a table cartogram should be easily recognized by readers, however, no methods have focused to enhance the readability. This study proposes a method that defines table cartogram generation as an optimization problem and attempts to minimize vertical and horizontal deformation. Since the original tables are comprised of regular quadrangles, this study uses quadrangles to express cells in a table cartogram and fixes the outer border to attempt to retain the shape of a standard table.
This study proposes a two-step approach for table cartogram generation with cells that begin as squares and with fixed outer table borders. The first step only adjusts the vertical and horizontal borders of cells to express weights to the greatest possible degree. All cells maintain their rectangular shape after this step, although the limited degree of freedom of this operation results in low data representation accuracy. The second step adapts the cells of the low-accuracy table cartogram to accurately fit area to weight by relaxing the constraints on the directions of borders of cells. This study utilizes an area cartogram generation method proposed by Inoue and Shimizu (2006), which defines area cartogram generation as an optimization problem. The formulation with vertex coordinate parameters consists of an objective function that minimizes the difference between the given data and size of each cell, and a regularization term that controls the changes of bearing angles. It is formulated as non-linear least squares, and is solved through the iteration of linear least squares by linearizing the problem at the coordinates of vertices and updating the estimated coordinates until the value of the objective function becomes small enough.