Learning from Users' Interactions with Visual Analytics Systems.
Abstract: Experts in
disparate fields from biology to business are increasingly called upon to make decisions
based on data, but their background is not in data science, which is itself a separate
field requiring years to master. Machine learning approaches tend to focus on finding a
black-box answer, which the user may not understand or trust. Visualization on its own
can leverage the power ... read moreof human insight, but may miss out on the computational power
available with automated analysis. Visual analytics researchers aim to provide tools for
domain experts to find the patterns they need in their data, and have recently been
interested in systems that combine the two approaches. One promising method is to blend
the best of visualization and machine learning by building systems that provide
interfaces for users to explore their data interactively with visual tools, gather their
feedback through interaction mechanisms, and apply that feedback by using machine
learning to build analytical models. In this dissertation, I discuss my research on such
systems, showing techniques for learning from user interactions about the data and about
the users themselves. Specifically, I first describe a prototype system for learning
distance functions from user interactions with high-dimensional data. These distance
functions are weighted Euclidean functions that are human-readable as the relative
importance of the dimensions of the data. Observing that users of such systems may be
required to review large amounts of data to be effective, I propose an algorithm for
better leveraging user efforts in this interactive context. Next, I show an adaptation
of the interactive learning prototype for text documents, with a study showing how to
make use of the vector representation of the distance functions for numerically
examining the analysis processes of the participants. Turning the focus of the learning
back onto the user, I provide a proof-of-concept that shows how models of users as
opposed to data can be learned from user interactions. Finally, I introduce the sketch
of a framework for future systems that will empower data stakeholders to find the
answers they need without leaving their comfort
Thesis (Ph.D.)--Tufts University, 2015.
Submitted to the Dept. of Computer Science.
Advisor: Remco Chang.
Committee: Robert Jacob, Greg Crane, Misha Kilmer, and Chris North.
Keyword: Computer science.read less