Role of human perception in cluster-based visual analysis of multidimensional data projections

Visualization of high-dimensional data requires a mapping to a visual space. Whenever the goal is to preserve similarity relations, multidimensional projections or other dimension reduction techniques are commonly used to project high-dimensional data point to a 2D point using a certain strategy for the 2D layout.Typical analysis tasks for projected multidimensional data do not necessarily match the expectations of human perception. Learning more about the effectiveness of projection layouts from a users perspective is an important step towards consolidating their role in supporting visual analytics tasks. Those tasks often involve detecting and correlating clusters. To understand the role of orientation and cluster properties of size, shape and density, we first conducted a study with synthetic 2D scatter plots, where we can set the respective properties manually. Then we picked five projection methods representative of different approaches to generate layouts of high dimensional data for two domains, image and document data. The users were asked to identify the clusters on real-world data and answers to questions were compared for correctness against ground truth computed directly from the data. Our results offer interesting insight on the use of projection layouts in data visualization tasks.