Database Reference
In-Depth Information
for each specific feature (e.g. 'right-turn speed') to provide a multi-
resolution view of the data. Lastly, to utilize this feature hierarchy, the
authors propose classification using hierarchical prediction rules (CHIP)
which is based on FOIL [70]. CHIP learns a set of rules by exploring
features in a top down manner, such that it will first attempt to classify
based on high level features in the hierarchy. To determine if it should
expand a lower feature level, CHIP computes the information gain from
using the higher resolution features. The experiments show that the
proposed technique significantly outperforms a basic SVM classification.
In fact, ROAM is shown to consistently outperform competing methods
as both the number of trajectories and the number of motifs increases.
Combining several of these ideas, [8, 50] build high-order models of
user movements in which subtrajectories may be identified as complet-
ing specific tasks. By high-order modeling, we refer to the ability of a
model to assign semantically meaningful labels to segments of a trajec-
tory (e.g. “going to work”) and include more abstract attributes (e.g.
mode of transportation) over the raw trajectory data. In this work, the
objective is to be able to answer queries not about the specific trajec-
tory, but about the purpose of the movement (e.g. given a trajectory is
it more likely that the subject is going to work or the grocery store?).
Specifically, Liao et al. [50] introduce a dynamic Bayesian network model
which incorporates high level goals, such as “going to work” or “going
to the supermarket” into the model which may be inferred from low
level location data (e.g. GPS). The result is a model which is capa-
ble of querying a variety of aspects of a user's movement. The authors
show that their proposed model can be learned in a completely unsuper-
vised manner, though applying the semantic categories such as “going
to work” must be supervised, and is able to identify locations of interest
and abnormal behavior in a real data trace.
In addition to mining patterns of movement , a related problem is to
identify those locations that are visited by a large number of trajectories.
Using the abundant amount of GPS trace data available over a region,
it is possible to find specific locations that are of interest (e.g. Statue of
Liberty, good restaurants, popular bars, etc.) [9, 84, 102]. For instance,
Zheng et al. [102] develop a PageRank-like algorithm for mining interest-
ing locations by considering each user's travel experience. The basic idea
is that users that are well traveled within a region of interest will likely
know more of the relevant locations and thus a visit from one of these
travel authorities should be weighted more than a visit from a tourist
who does not know the local area well. Alternatively, Uddin et al. [84]
identify regions of interest (ROIs) from trajectory data by looking for
dense areas (at least N mobile objects in a fixed area) in which mobile
Search WWH ::




Custom Search