Human in the Loop: Iterative, Interactive Visual Model Refinement

• Activity Number: 678
• Type: Topic Contributed
• Date/Time: Thursday, August 8, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Eli T. Brown*+
• Companies: Tufts
• Keywords: visual analytics ; visualization ; interactive ; user model ; provenance ; metric learning

Abstract

Visual analytics has leveraged visual representation to help users explore complex relationships, and machine learning and statistics have explored model-learning algorithms that take advantage of some supervision from humans. However, where the former lacks emphasis on model building, the latter lacks real-time user involvement. In this talk, I will present a prototype system published in VAST 2012 that provides a projection of data and allows a human user to interact directly with the points. Through several iterations the user improves the projection to agree with expert knowledge, and we learn a corresponding distance function over the data. We use a weighted Euclidean distance function, so the result is human-readable and provides an importance weight for each dimension of the data that can be considered a quantification of the user's understanding of the data. Aside from the existing system, I will discuss some of the challenges and opportunities in learning from humans on high-dimensional data.