A Framework for Data Warehousing and Mining in Sensor Stream Application Domains - Evolving Application Domains of Data Warehousing and Mining - page 114

useful and understandable patterns for the end
users to perform data analysis. Many research
projects have been conducted by different orga-
nizations regarding wireless sensor networks;
however, few of them discuss the sensor stream
processing infrastructure, and the data warehous-
ing and data mining issues need to be addressed in
the sensor network application domains. There is
a need for new methodologies in order to extract
interesting patterns in a sensor stream application
domain. Since the semantics of sensor stream data
is application dependent, the extraction of interest-
ing, novel, and useful patterns from stream data
applications becomes domain dependent.
Some data warehousing and data mining
methods have been recently proposed to mine
stream data, for example in (Manku 2002, Chang
2003, Li 2004, Yang 2004, Yu 2004, Dang 2007),
the authors proposed algorithms to find frequent
patterns over the entire history of data streams.
In (Giannella 2003, Chang 2004, Lin 2005, Koh
2006, Mozafari 2008), the authors use different
sliding window models to find recently frequent
patterns in data streams. These algorithms focus
on mining frequent patterns with one scan over
the entire data stream.
In (Chi, 2004), Chi et al considers the problem
of mining closed frequent itemsets over a data
stream sliding window in the Moment algorithm,
and in (Li, 2006), the authors proposed the New-
Moment algorithm which uses a bit-sequence
representation of items to reduce the time and
memory needed. The CFI-Stream algorithm in
(Jiang, 2006) directly computes the closed itemses
online and incrementally without the help of any
support information. In (Li, 2008), Li et al pro-
posed to improve the CFI-stream algorithm with
bitmap coding named CLIMB (Closed Itemset
Mining with Bitmap) over data stream's sliding
window to reduce the memory cost.
Besides pattern mining in data stream applica-
tions, as the number of data streaming applications
grows, there is also an increasing need to perform
association mining in data streams. One example

application is to estimate missing data in sensor
networks (Halatchev, 2005). Another example
application is to predict frequency of Internet
packet streams (Demaine, 2002). In the MAIDS
project (Cai, 2004), an association mining tech-
nique is used to find alarming incidents from data
streams. Association mining can also be applied
to monitor manufacturing flows (Kargupta, 2004)
to predict failures or generate reports based on
accumulated web log streams. In (Yang, 2004),
(Halatchev, 2005), and (Shin, 2007), the authors
proposed using two, three, and multiple frequent
pattern based methods to perform association
rule mining.
In general, these approaches have focused on
mining patterns and associations in data streams,
without considering an application domain. As
a consequence, these methods tend to discover
general patterns, which for specific applications
can be useless and uninteresting. Stream patterns
are usually extracted based on the concept of
pattern frequency. With no semantic or domain
information, the discovered patterns cannot be
applied directly to a specific domain.
In this topic chapter, we present a data ware-
housing and mining framework where the users
give to the data the semantics that is relevant for
the application, and therefore the discovered pat-
terns will refer to a specific domain. We will also
discuss the issues needed to be considered in the
data warehousing and mining components of this
framework for sensor stream applications.
The remaining of the chapter is organized as
follows: in the background section, we present
some basic concepts about sensor data applica-
tions and the data warehousing and mining issues
needed to be considered in stream data applica-
tions. Following with the background information,
we present a framework for domain-driven data
warehousing and mining from sensor streams.
The case study section shows the experimental
results with real data in two application domains.
Finally, we conclude the chapter and discuss the
future trends.