Trajectories of Search - Mapping Scientific Frontiers: The Quest for Knowledge Visualization - page 155

locates a relevant document, he tends to explore documents in the immediate
neighboring area, just as we expected. The frequency of long-range jumps across
the space decreased as the user became familiar with the structure of the space. The
trajectory eventually settled to some fine-grained local search within an area where
the majority relevant documents are placed, and it didn't move away from that area
ever since, which was also what we expected.
In the trajectory replay, the time spent on a document is animated as the radius
of a green disc growing outward from where the document is located. This design
allows us to find out whether the majority of large green discs appear in areas with a
high density of relevant documents, and whether areas with a low density of relevant
documents will only have sporadic passing navigation trails.
We found that users were able to mark certain documents extremely fast. For
example, user jbr apparently spent almost no time to determine the relevancy of
documents 80, 20, and 64 and marked them in blue. It seems once users have
identified two relevant documents, they tend to identify relevant documents in
between very quickly. Explicit links in the visualization play a crucial role in
guiding the course of navigation of users. Not only users follow these links in their
navigation, but also make their relevance judgment based on the cues provided by
these visible links. In other words, users have relied on these explicit links to a
considerable extent when they assess the profitability of a document.
Trajectory maps are designed so that an outline of the trajectory from the previous
task can be preserved and carried over to the next task. If a user spends a long time at
a document in task A, the accumulative trajectory map starts with this information.
We expected to see a user would gradually narrow down the scope of active search
areas. In addition, as users become increasingly familiar with the structure and
content of the underlying thematic space, there would be no need for them to re-
visit areas with low profitability.
Figure 4.10 shows the “Zoom in” stage of the search. The search trail never went
to the area identified in the immediately previous “Overview first” stage. The next
stage, “Details on demand,” is shown in Fig. 4.11 .
Figure 4.12 shows the trajectories of the same user jbr for four tasks. These maps
reveal that the user spent longer and longer time in areas with relevant documents.
In the last trajectory map for task D, the user began to forage information in new
areas.
Trajectories of individual users have revealed many insightful findings. The next
step is to extract behavioral patterns from the group of users as a whole. From a
social navigation point of view, not only one has to understand the characteristics of
the trajectory of individual users in a spatial-semantic space, but also to identify the
commonality across individuals' behavioral patterns.
Hidden Markov Models allow us to describe and predict sequential behavior
characteristics of users foraging information in thematic spaces. We categorize
users' information foraging actions into three types of action events:
Node over
Node click, and
Node mark.