Bayesian Geographical Multi-Dimensional Scaling

Abstract. Multi-dimensional scaling (MDS) is a popular method of visualizing the similarity of individuals in a dataset. When dissimilarities between individuals in a dataset are measured, MDS projects these individuals into the (typically two- or three-dimensional) map. In this map, because similar individuals are projected to be close to one another, distances between individuals correspond to their dissimilarities. In other words, MDS makes a similarity map of a dataset.

Some of the dissimilarities and distances have a strong relation to the geographical location. For example, time distances are similar to geographical distances, and regional features will be similar if the regions are close together. Therefore, it will be useful to compare the MDS projection and geographical locations. However, because MDS projection is not concerned with the rotation, parallel translation, and similarity expansion, it might be difficult to compare the projection to the actual geographical locations. When geographically related similarities are visualized, projected locations should be bound to the geographical locations.

In this article, we propose Bayesian Geographical Multidimensional Scaling (BGMDS), in which geographical restrictions of projections are given from a statistical point of view. BGMDS gives not only geographically bound projections, but also incorporates the uncertainty of the projections.