Database Reference
In-Depth Information
ommending specific non-spatial products or items on the basis of spatial
history, as discussed in [101].
This general principle can also be applied for geographical topic dis-
covery and comparison from GPS-associated documents [172]. While
topic modeling of documents is widely known, the use of geographic
information in the process provides rich opportunities for adding addi-
tional insights into the process. Many interesting concepts, including
cultures, scenes, and product sales, correspond to specialized geographi-
cal distributions. The goal of geographical topic discovery is to discover
such interesting concepts. The two main questions in this context are as
follows:
What are the coherent topics of interest in the different geograph-
ical regions?
How can the different topics be compared across the different ge-
ographical regions?
The work in [172] proposes and compares three different models which
use pure location, pure text, and a joint model of location and text,
which is referred to as LGTA (Latent Geographical Topic Analysis).
The approach is used on several data sets from the
Flickr
web site. It
is shown that the first two methods work in some data sets but fail in
others, whereas LGTA works well in all data sets at finding regions of
interest and also providing effective comparisons of the topics across dif-
ferent locations. This suggests that geographical data and content data
provide complimentary information to one another for the mining pro-
cess. Further work along this direction in the context of
topic evolution
is proposed in [169]. From a real-time perspective, it is often useful to
utilize location information for providing context-sensitive newsfeeds to
users [19].
An interesting application in [32] shows that the latent information
in user trajectories, which are extracted from the GPS data in photos
can even be used to generate travel itineraries. For example, the media
sharing site
Flickr
, allows photos to be stamped by the time of when
they were taken and be mapped to points of interest with the use of
geographical and tag meta-data. This information can be used to con-
struct itineraries with a two-step approach. First, the photo streams of
individual users are extracted. Each photo stream provides estimates on
where the user was, how long he stayed at each place, and what was the
transit time between places. In the second step, all user photo streams
are aggregated into a Point of Interest (POI) graph. Itineraries are then
automatically constructed from the graph based on the popularity of the
POIs, and subject to the user time and destination constraints.
