Geography Reference
In-Depth Information
1. The automatic approaches usually utilize different constraints and assumptions
that assure adequate performance. Such assumptions are for example the
following:
a. The representative tags or events are determined by the semantics of the
textual information and not by the geographical constraints. This enables
finding only one event per area (cluster). Clearly, if there are several events
occurring in one place at different time or at the same time (overlapping
events), only the most significant one will be selected. The significance of
the events is purely algorithm dependent and can lead to loss of information
about other events.
b. The significance of the place is usually determined by the number of photos,
users or other heuristics. Thus, the event with a few number of photos or
people can be missed. For example, Jaffe et al. (
2006
) note that more than
1,000 photos on a city scale are required in order to obtain meaningful
results.
c. The significance of the place is determined by the uniqueness of the textual
semantics within the cluster. It means that in order to find a significant event
in a cluster, other clusters, surrounding that cluster, should be analyzed and
semantics extracted from them to be taken into consideration. This makes
the algorithms non-scalable when large areas are used for exploration.
2. Experiments are performed using clean-room data samples, where class labels
are manually prepared or taken from existing benchmark sources. Therefore,
such issues as geographical errors, different languages or mistakes made in
textual information are usually not raised.
3. Different representation models as well different algorithms produce different
results.
4. Algorithm accuracy is reported with respect to the best-tuned parameters
applicable to the training data. No real experiments were performed on arbi-
trary data.
In addition, all the mentioned approaches are user centric, aiming at providing
solutions for exploration but not for analysis. Examples are: the representative tag
viewer by Kennedy et al. (
2007
), tag maps by Jaffe et al. (
2006
) or tag mapping
''world explorer'' by Ahern et al. (
2009
). In contrast to these approaches, our paper
aims at the analysis of places and events.
We claim, however, that combining the above mentioned techniques with
geospatial visual analytics methods, GeoComputation, spatial and spatio-temporal
data mining create new opportunities for the analysis of spatio-temporal data. The
most important difference between the existing approaches and the methodology
proposed by us, is the way in which event clusters are obtained. In contrast to the
semantic-centric approach, we use spatio-temporal clustering based on geo-
graphical properties of the data as commonly used in geographical analytics. This
allows us to apply different techniques like time-series, text or multimedia analysis
