Information Retrieval Perspective to Interactive Data Visualization

Abstract

Dimensionality reduction for data visualization has recently been formulated as an information retrieval task with a well-defined objective function. The formulation was based on preserving similarity relationships defined by a metric in the input space, and explicitly revealed the need for a tradeoff between avoiding false neighbors and missing neighbors on the low-dimensional display. In the harder case when the metric is not known, the similarity relationships need to come from the user. We formulate interactive visualization as information retrieval under uncertainty about the true similarities, which depend on the user's tacit knowledge and interests in the data. During the interaction the user points out misses and false positives on the display; based on the feedback the metric is gradually learned and the display converges to visualizing similarity relationships that correspond to the tacit knowledge of the user.