Earl Rennison
Visible Language Workshop
MIT Media Lab
20 Ames Street
Cambridge, MA 02139
Each information space is multidimensional in that it can change scale, orientation, perspective,
representation and presentation as the user navigates through the space, as illustrated in Figures
1a, 1b, and 1c. Because the visual space represents the complex relationships of the information
objects, it is highly abstract and does not resemble anything we see in the physical world. Rather,
the fluid non-linear movement through the space more closely resembles our unconscious and is
designed to convey the associative properties of memory and thought.
Figures 1a, 1b and 1c:
Sequence, from left to right, of a progressive zoom into the "weather" cluster of news information
objects. The first image shows a wide angle view of multiple clusters. As the user zooms closer
(second image), the system reveals headlines of articles contained in the cluster. Further zooming
(third image) reveals the body of the articles and, in some cases, images, video, or sound
clips.
Figure 1a (Caution: Large 71465 mbyte GIF file)
Each information space is dynamic and responsive to user from information objects has implicit
meaning and forms queries or requests for more details, more generalization, different
perspectives, and so forth. This interaction forms a dialog that is formally defined by a visual
discourse grammar. As a result, the user not only browses "system output" by navigation,
but also expresses "input commands" in the same manner. The advantage of this approach is that
the user is always focused on the information content and does not have to deal directly with the
syntax of the dialog. An example of an interaction with an information space is illustrated in Figure
1a, 1b, and 1c.
Our approach to this problem is to learn what information a user prefers and use these
preferences to construct an operator that transforms a non-bias information space into a
personalized one to assist with directed searches. This approach is desirable because it does not
require the user to explicitly specify or characterize the information he or she is searching for and
because it also allows the user to browse while searching.
The implementation of this approach relies on four key elements:
The system employs two levels of learning. First, the system learns the relationships between
documents contained within an information base. Second, the system learns a user's information
interests as he or she navigates through the information space and reads the documents. When
the user reads a document, the system passively inserts symbol associations into the user
interest model. Additionally, the user can actively reinforce a preference or dislike for a particular
document or part of a document. A user can control the user preference learning process by
adjusting the learning rate and an aging or forgetting rate.
The learned user interest model is used to construct an operator that can selectively
transform a non-biased information space into a personalized information space. When the
personalization transformation is applied to an information space, the system animates from one
visual representation to the next to aid the user in understanding the changes.
The biasing process in the Personalized Galaxies of Information imposes a personalized
organization structure on the information. This organization is not explicitly specified by the user.
Rather, it is indirectly learned by the systeabout the information, the system learns about the user.
As this process is repeated, both the user and the system become more efficient. This approach
is advantageous in environments such as the internet and World Wide Web where information is
constantly evolving and changing.
Abstract
The Personalized Galaxies of Information demonstration presents a new interface approach for
visualizing, navigating and accessing information objects in a large body of unstructured
information, such as on-line news stories, photographs a! electronic mail; and World Wide Web
documents. The system provides mechanisms to analyze the relationships between information
objects and builds a representation of the underlying structure of the entire body of information.
This relational structure is used to construct a visual information space with which the user
interacts to explore the contents of the information base. The system also uses a learning
algorithm to adaptively customize the presentation of information to a particular user's interests.
This dynamic, personalized structuring of information helps users perform directed searches while
simultaneously affording general browsing in a fluid and seamless environment.
Keywords:
Information visualization, abstracted information spaces, 3D interactive
graphics, user interest models, reinforcement learning.
Introduction
The rapid expansion of the internet and on-line services has led to the availability of an ever-
increasing amount of information!line services are often authored independent of other related
documents. The process of linking documents together has historically been the job of an editor.
However, manual editing of information is becoming increasingly more difficult due to the scale
and complexity that is emerging in our globally interconnected community. As a result, we need an
infrastructure that automatically builds correlations and relationships between information objects
(hence linking the objects together), and provides coupled dynamically constructed information
spaces and facilitates interactive navigation and intuitive access to related or correlated
information. With such an environment, we would like to provide the ability to simultaneously
search and browse. Hence, a system must adapt to different users and must also follow the
changing needs of individual users. This demonstration illustrates our approach to these
problems.
VISUALIZATION AND INTERACTION WITH ABSTRACTED INFORMATION
SPACES
The Personalized Galaxies of Information demonstration illustrates a new metaphor for
information visualization: abstracted information spaces [3]. These abstracted
information spaces have three important properties:
The structure of each information space emerges from a collection of information objects. The
structure is emergent in that it is not defined by any single document, nor authored by a single
author. The structure of each information space emerges from a collection of information objects.
The structure is emergent in that it is not defined by any single document, nor authored by a single
author. Rather the structure emerges from the inherent relationships within a collection of
documents.
Figure 1b (Caution: Large 119128 mbyte GIF file)
Figure 1c (Caution: Large 341745 mbyte GIF file)
PERSONALIZED INFORMATION VISUALIZATION
The Galaxies of Information visualization approach lends itself to browsing of information,
providing an intuitive interface for exploration. However, one inherent problem with this approach
is that the structure of the space changes as the information changes. As a result, the location of
information also changes, making it difficult to find information that is of specific interest to a user,
as in a directed search. Hence, what is needed is the ability to reorganize the information space to
meet the personal needs of the user.
The user preferences, or user interest model, in this approach are represented using an
associative relation network (ARN), as described in [1]. vector space (as used in Latent Semantic
Indexing[2]) and semantic networks[1]. This user interest model is used to directly bias the
underlying structure of the information to reflect the way a user associates information.
Acknowledgments
The author would like to acknowledge the special contribution of the late Muriel Cooper who had a
significant impact on the direction of this research. The author would Workshop. This work was
sponsored by ARPA, JNIDS, NYNEX and Alenia.
