Observation-Level Interaction with Clustering and Dimension Reduction Algorithms

Abstract

Observation-Level Interaction (OLI) is a sensemaking technique relying upon the interactive semantic exploration of data. By manipulating data items within a visualization, users provide feedback to an underlying mathematical model that projects multidimensional data into a meaningful two-dimensional representation. In this work, we propose, implement, and evaluate an OLI model which explicitly defines clusters within this data projection. These clusters provide targets against which data values can be manipulated. The result is a cooperative framework in which the layout of the data affects the clusters, while user-driven interactions with the clusters affect the layout of the data points. Additionally, this model addresses the OLI "with respect to what" problem by providing a clear set of clusters against which interaction targets are judged and computed.