Pollux: Interactive Cluster-First Projections of High-Dimensional Data

Semantic interaction is a technique relying upon the interactive semantic exploration of data. When an analyst manipulates data items within a visualization, an underlying model learns from the intent underlying these interactions, updating the parameters of the model controlling the visualization. In this work, we propose, implement, and evaluate a model which defines clusters within this data projection, then projects these clusters into a two-dimensional space using a “proximity≈similarity” metaphor. These clusters act as targets against which data values can be manipulated, providing explicit user-driven cluster membership assignments to train the underlying models. Using this cluster-first approach can improve the speed and efficiency of laying out a projection of high-dimensional data, with the tradeoff of distorting the global projection space.