Database Reference
In-Depth Information
A closely related problem is that of mining frequent episodes in sequences [
55
,
56
,
95
,
94
,
144
]. While this problem is not the same
1
as frequent or sequential pattern
mining, it is very closely related, and often uses similar frameworks which use
support in order to quantify the significance of the underlying patterns. Other related
problems include that of periodic pattern mining in which patterns are constructed on
the basis of the seasonality in time-series sequences [
43
,
58
,
92
]. Such patterns are
often useful for clinic diagnosis in time-series patterns such as ECG measurements.
Sequence data provides a much richer domain than non-sequential data for mining
purposes. A method known as MARBLES proposes methods for finding association
rules between episodes [
38
]. The problem of mining train delays with the use of
sequence mining is discussed in [
39
]. Such methods are useful for finding how
different episodes are related to one another.
10
Spatial and Spatiotemporal Applications
With advances in mobile sensing technology, an important emerging scenario is of
social sensing [
3
]. In this case, the data is collected from mobile phones continuously
over time, and much of this data is in the form of GPS-based location data. GPS-based
location data can also be used in order to construct trajectories. In many cases, it is
desirable to determine clusters and frequent patterns from the underlying trajectories.
Frequent pattern mining methods have frequently been used for clustering spa-
tiotemporal data. An example of such a technique is the
Swarm
method proposed in
[
84
]. In this approach, the data is first pre-processed into different snapshots. In each
snapshot, a discrete value is used to indicate the cluster membership of an object. For
example, clustering could be applied to each snapshot in order to obtain a discrete
value for the cluster membership. Objects which have the same discrete identifier
over multiple snapshots clearly correspond to a SWARM which moves together.
Therefore, the approach in [
84
] defines a pattern-based model in which frequently
occurring sequences of discrete values are reported together with the objects, which
correspond to such sequences. Other related models for pattern mining in spatiotem-
poral data are discussed in [
22
,
52
,
53
]. Many of these models can benefit from the
use of frequent or sequential pattern mining methods.
Methods have also been designed for performing classification from trajectory
data with the use of frequent-pattern mining methods. In particular, the method
in [
82
] determines important patterns which are related to the rare classes. These
patterns are then used in order to predict the rare class. Thus, this approach can
be used for supervised anomaly detection in spatiotemporal data. Another method
proposed in [
81
] finds movement fragment patterns by spatial overlay. These are then
used in order to identify outliers with the use of pattern-based classification.
1
Many other kinds of methods such as Markov Models [
55
] are used in order to solve this problem.
