How to evaluate a subspace visual projection in interactive visual systems? A position paper

This paper presents a position paper on subspace projection evaluation methods in interactive visual systems. We focus on how to evaluate real information rendered through the visual data projection for the mining of high dimensional data sets. To do this, we investigate automatic techniques that select the best visual projection and we discuss how they evaluate the projections to help the user before interactivity. When we deal with high dimensional data sets, the number of potential projections exceeds the limit of human interpretation. To find the optimal subspace representation, there are two possibilities, the first one is to find the optimal subspace which reproduces what really exists in the original data: getting the existing clusters and/or outliers in the projection. The second possibility consists in researching subspaces according to the knowledge discovery process: discovering novel, but meaningful information, such as clusters and/or outliers from the projection. The problem is that visual projection cannot be in adequation with the subspaces. In some cases, the visual projection can show some things that do not really exist in the original data space (which can be considered as an artifact). The mapping between the visual structure and the real data structure is as important as the efficiency and accuracy of the visualization. We examine and discuss the literature of Information visualization, Visual analytic, High dimensional data visualization, and interactive data mining and machine learning communities, on how to evaluate the faithfulness of the visual projection information.