One first step to get insights about a dataset can be its visualization using dimensionality reduction (DR). However, DR processes induce a loss of information that needs to be quantified in order to evaluate the quality of their results. Furthermore, two DR visualizations with a similar loss value can be really different in the eyes of the user. This paper presents DR quality measures developed in the machine learning community, as well as visual quality measures considered in the information visualization community, which can be used to assess interpretability. We propose to combine several measures from these two categories in order to be able to predict and study users' understanding of DR visualizations.
|titre||SafeML ICLR Workshop|
|Lieu de publication||New Orleans, Louisiana|
|Etat de la publication||Publié - 2019|