Trajectories of Search - Mapping Scientific Frontiers: The Quest for Knowledge Visualization - page 149

spatial-semantic interfaces. A significant number of overview maps are designed
in this way. Second, it contains task models of visual navigation. We refer to
Shneiderman's information visualization taxonomy (Shneiderman 1998 ). In the
following example, four thematic topics were chosen to form the basis of an exper-
imental design. Information visualization techniques such as Pathfinder networks
and minimum spanning trees (MST) were used in the experiment. Hidden Markov
Models (HMMs) can be derived from users' navigation sequences recorded in each
session. Each user' performance is measured in terms of recall and precision, as
traditionally done in information retrieval. Additional data, in particular, the course
of navigation, were also collected. The navigation trails of the most successful users
were used to feed in the HMM modeling process. Finally, synthesized user trails
are generated from the HMMs and animated within the corresponding thematic
spaces.
In order to study users' information foraging behavior, we constructed four
thematic spaces based on news articles from Los Angeles Times retrieved from
the Text Retrieval Conference (TREC) test data. Each thematic space contains the
top 200 news articles retrieved through a single keyword query to the document
collection. The four keywords used were alcohol, endangered, game, and storm.
Corresponding spaces were named accordingly by these keywords.
For each thematic space, we generated a document-to-document similarity
matrix using Latent Semantic Indexing (LSI) (Chen and Czerwinski 1998 ; Deer-
wester et al. 1990 ). A few types of spatial-semantic interfaces were produced,
including Pathfinder networks (PF) and minimum spanning trees (MST). In MST-
based visualization, N-1 explicit links connect all the documents together. Users
can see these links on their screen. In PF-based visualization, additional explicit
links are allowed as long as the triangular inequality condition is satisfied. In our
examples, the PF version tends to have only a handful of extra links at most in
comparison to the MST version. Detailed descriptions of the use of these techniques
for information visualization can be found in (Chen 1999a ).
Sixteen users, nine male and seven female students from a British university,
participated the experiment. They performed a series of tasks in each thematic space
through spatial-semantic interfaces. Usage data was logged to a computer file in
each session, including the occurrence of an event, the time stamp of the event, and
the target document on which the event takes place.
The design of the tasks follows Shneiderman's mantra: overview, zoom, filter,
details on demand, which highlights users' cognitive needs at various strategic
stages in visual information retrieval. At the top level with Task A, users need to
locate and mark documents relevant to a search topic. For example, in the Alcohol
space, users were asked to locate and mark any documents that mention an incident
of drink driving. Twenty to twenty five documents were judged as relevant by
experts for TREC conferences. Task B is more specific than Task A, and so on.
We expect that users would narrow down the scope of their search from Task A
through Task D and that this should be evident in their trails of navigation.