Trajectories of Search - Mapping Scientific Frontiers: The Quest for Knowledge Visualization - page 152

o k . d i /
d i

b ik D

The reason we choose the profitability function of a document space as the state
space and the three types of events as the observation symbols is based on the
consideration that the sequence of activities such as node over, node click, and node
mark is a stochastic process. This observable process is the function of a latent
stochastic process - the process of estimating the profitability of documents in the
thematic space by a user because which document the user will move to in his/her
next step is very much opaque to observers.
We constructed HMMs based on the actual trails recorded from sessions of the
experiment. HMMs are both descriptive and normative - not only can one describe
what happened with information foraging sessions, but also can one predict what
might happen in similar situations. HMMs provide insights into how users would
behave as they are exposed to the same type of structural and navigational cues in
the same thematic space.
We defined the basic problems as follows. The first basic question states that
given observation O D (o1, o2, :::, oT), which is a sequence of information
foraging actions of a user, and model œ D (A, B, ), efficiently compute P(O j œ).
Given two models œ1andœ2, this can be used to choose the better one. We first
derived an HMM model from the log files of two users: one has the best performance
score, but without any node click events; the other has all types of events. This model
is denoted as œ log . Given an observation sequence, it is possible to estimate model
parameters œ D (A, B, ) that maximize P(O j œ), denoted as œ seq . The navigation
sequence of the most successful user provided the input to the modeling process.
The second basic question states that given observation O D (o1, o2, :::,oT)
and model œ find the optimal state sequence q D (q1, q2, :::, qT). In this case,
we submited the navigation sequences of users to the model œlog and animated the
optimal state sequences within the thematic space. In this way, we can compare the
prevalent navigation strategies. Such animation will provide additional navigational
cues to other users.
Finally, the third basic question states that given observation O D (o1, o2,
:::, oT), estimate model parameters œ D (A, B, ) that maximize P(O j œ). We
focused on the most successful user in searching a given thematic space. If a user
is clicking and marking documents frequently, it is likely that the user has found a
high profitable set of documents.

4.1.4.2 Visualizing Trails of Foraging

Figure 4.7 is an annotated screenshot of the graphical interface design, which
explains how users' navigation sequences are animated. Documents in red are not
relevant to the search tasks. The course of navigation appears as dotted yellow
links. Relevancy judgments made by experts are provided in the TREC test data.
Documents relevant to the original search are marked with a bright yellow dot in the