Computational Aspects: How Landmarks Can Be Observed, Stored, and Analysed - Landmarks: GIScience for Intelligent Services - page 164

5.5.2 Taxonomies

Given the previous discussion, using types rather than individuals seems to be the
more promising approach. Here, properties of individual geographic objects do not
matter as they are inferred by some heuristics. These heuristics make use of a general
assessment of a specific type's suitability as a landmark. For example, it may be
argued that in general a pub is more salient than a doctoral clinic. Consequently,
much less data would be required to make such an approach work. In fact, a POI
database that categorizes its entries according to type would suffice. This is the
approach taken by Duckham et al. [ 9 ] , which has been discussed earlier in this
chapter.
Götze and Boye [ 16 ] recently suggested using a machine learning approach to
determine suitable landmark candidates. In their approach, landmark candidates
are described using a feature vector that may contain data, such as distance from
the route, but also categorical information, for example, whether the candidate
is a restaurant. Their approach determines user preferences for landmarks from
route directions generated by the users themselves. Depending on which candidates
appear in these descriptions, preferences for specific kinds of landmarks are learned
by the system and offered in system-generated descriptions in the future. This takes
away the necessity for a detailed data collection for every landmark candidate since
the feature vector uses fairly simple attributes. Still, these vectors need to be filled
and candidates need to be collected for the approach to work—and most importantly,
users need to be motivated to describe routes to themselves.
Even more, as argued above, existing POI databases usually exhibit a sparse and
uneven distribution of POIs across the space they cover. For example, Richter [ 38 ]
has shown for the WhereIs routing service that assuming an even distribution across
Australia—which is clearly not the case—there would only be one POI object
every 45 km 2 . Thus, clearly new ways of collecting a sufficient number of landmark
candidates with sufficient detail are required. These will be discussed in the final
section of this chapter.

5.5.3 Crowdsourcing as an Alternative Approach

Crowdsourcing [ 46 ] is an approach to acquire desired content or services from
a large number of people ('the crowd') rather than from traditional suppliers
('the individual'). It is a predominant approach in open source software develop-
ment, where a (large) number of people interact and co-develop a software project.
When it comes to collecting potential landmark candidates we may use similar
approaches as in crowdsourcing. Since our interest is rather in data (or information)
than services, we may rather speak of user-generated content [ 27 ] or volunteered
geographic information [ 15 ] . These three terms are not synonymous, however for
the purposes of the following discussion the differences between them are not