Towards Supporting Interpretability of Clustering Results with Uncertainty Visualization

Abstract

Interpretation of machine learning results is a major challenge for non-technical experts, with visualization being a common approach to support this process. For instance, interpretation of clustering results is usually based on scatterplots that provide information about cluster characteristics implicitly through the relative location of objects. However, the locations and distances tend to be distorted because of artifacts stemming from dimensionality reduction. This makes interpretation of clusters difficult and may lead to distrust in the system. Most existing approaches that counter this drawback explain the distances in the scatterplot (e.g., error visualization) to foster the interpretability of implicit information. Instead, we suggest explicit visualization of the uncertainty related to the information needed for interpretation, specifically the uncertain membership of each object to its cluster. In our approach, we place objects on a grid, and add a continuous ''topography'' in the background, expressing the distribution of uncertainty over all clusters. We motivate our approach from a use case in which we visualize research projects, clustered by topics extracted from scientific abstracts. We hypothesize that uncertainty visualization can increase trust in the system, which we specify as an emergent property of interaction with an interpretable system. We present a first prototype and outline possible procedures for evaluating if and how the uncertainty visualization approach affects interpretability and trust.