Database Reference
In-Depth Information
all analysis is performed only on the discrete representations of the underlying data,
since it is directly suited to the frequent pattern mining framework. It should be
pointed out that the work done on mining patterns from biological sequences can
also be applied directly to temporal sequences with a few modifications [
34
,
35
,
89
,
104
,
105
,
111
,
125
]. Another interesting similarity between biological and temporal
data is that (unlike other frequent pattern mining scenarios) biological data often
contains a small number of very long rows. As a result, row-enumeration techniques
are often used in these methods.
One of the most common applications of pattern analysis techniques in the tem-
poral context is that of event detection [
23
,
64
,
66
,
80
,
121
]. It should be noted that
event detection can be considered a temporal version of the classification problem,
in which labels are associated with time-stamps rather than records. The connection
between pattern mining and classification has already been discussed in a previous
section. Therefore, it is natural to utilize sequential pattern mining methods in the
context of rule-based methods. In these cases, the data consists of a set of sequences
defined on
base feature events
, and a
class event
which needs to be predicted from the
patterns in the sequences defined by the feature events. The idea in most such tech-
niques is that events can be predicted by particular sequences in the underlying data.
This is used to construct temporal classification rules, which correspond to events.
Such temporal classification rules will typically contain a sequence of feature events
on the left hand side and a class event on the right hand side. In addition, the rule
may contain a numerical lag value associated with it, which indicates the time lag
with which the event will occur after a particular sequence of feature events. Once
such rules have been mined, they are used for the prediction process, as in the case of
all pattern-based classifiers. Event detection with the use of frequent pattern mining
methods has been used frequently in the context of intrusion detection [
73
-
75
]. The
goal in these methods is to relate the temporal patterns of the features in the network
packets to the intrusion events. This model is then used in order to predict events. It
should be pointed out that these methods can be used more generally for a variety
of event detection problems beyond the intrusion scenario. An overview of classifi-
cation methods for sequential data with the use of rule-based methods is provided
in [
131
]. A closely related problem is that of mining process models from workflow
logs [
15
]. Sequence mining is also used in order to predict customer behavior in
telecommunications [
44
].
Time series data and sequence data are often mined for characteristic motifs.
Such motifs may often describe the important trends in the underlying data, and
can even be used for classification. An example of such an approach is discussed
in [
1
], where pattern-based rule mining is used to determine the class labels of the
underlying sequences. In this case, wavelet decomposition is applied to the sequences
in order to create a multi-granularity representation, in terms of which the rules can
be represented. The multi-granularity representation allows the construction of rules
which span different lengths of the time-series, as long as they are relevant to the
classification process. Note that this application is somewhat different from the event
detection problem, since labels are associated with individual time series, rather than
with specific instants in the time series.
