Integrating Preference Elicitation into Visualizations

Abstract

Modern technology has enabled researchers to collect large amounts of information in an expanding scope of research fields. At the same time, these new datasets are becoming more complex as evidenced in their increasing size and dimensionality. Managing and understanding these datasets has become a challenging problem. Visualizations attempt to address these concerns by creating meaningful graphical representations of data that can rapidly and accurately convey important information and interesting properties about the data to a researcher. However, many existing visualization algorithms are overwhelmed by the size of today's datasets. As a result, information is often forced off-screen due to a lack of visual resources. In previous work, we developed a navigation assistant to aid users with finding interesting data elements located off-screen. The assistant used a graph framework to provide way-finding cues and generate informative animated tours of the visualization. In order to identify which elements to include in this framework, the navigation assistant needs to model users' interests; i.e., their preferences.

The efficient collection and modeling of a user's preference information is a fundamental goal of preference elicitation. Many of these techniques have yet to be applied to real-world practical problems. We address the challenges of integrating a preference model and corresponding elicitation techniques into an environment not especially suited for collecting preference information, specifically, a visualization environment. Using combinations of explicit and implicit techniques, the navigation assistant collects preference information from users both before and during their interaction with a visualization. These techniques provide input to an underlying preference model used by the navigation assistant to dynamically add or remove elements from the graph framework. Using the preference model, the assistant attempts to create a description of a user's preferences, possibly revealing previously unknown interests.