DISTRIBUTED DATA MINING IN SENSOR NETWORKS - Managing and Mining Sensor Data - page 227

node determines outliers in its local dataset, and then broadcasts them to
other nodes for validation. The neighboring nodes repeat the procedure
until all of the sensor nodes in the network eventually agree on the
global outliers. This technique can be flexible with respect to multiple
existing distance-based outlier detection techniques. It has two major
advantages: (1) the outliers found this method are provably the same
that a centralized algorithm would find, and (2) the algorithm can easily
adopt to data and network changes. Because of these two advantages,
the technique is greatly suitable for WSNs. However, one drawback of
this method is that it requires a node to broadcast all the outliers to all
the other nodes for validation.
Zhang et al. [65] propose a distance-based technique to identify n
global outliers in continuous query processing applications of sensor net-
works. To overcome the broadcast issue of Branch et al. [10], [65] adopts
the structure of aggregation tree that do not require broadcasting of each
node in the network. Each node in the tree transmits some useful data
to its parent after collecting all the data sent from its children. The sink
then approximates the top n global outliers and sends these outliers to
all the nodes in the network for verification. If any node disagrees on
the global results, it will send extra data to the sink again for outlier
detection. This procedure is repeated until all the nodes in the network
agree on the global results calculated by the sink. A major drawback of
this technique is that it requires a tree topology to be overlaid on top of
the network and hence not suitable for any topology types.
4.3 Classification based approaches
Given examples of two kinds, an outlier detection problem can be
transformed to a classification problem. This trick has been widely ex-
plored in the data mining community and Chandola et al. [12] presents a
good overview on this topic. Even in the area of WSN, the classification
techniques that we have presented in Section 3 can be applied for outlier
detection in WSNs. One such example is the one-class support vector
machines algorithm that can learn a non-linear hyper surface via the
kernel trick. Rajasegarar et al. [48] use this model for outlier detection.
In the first phase of this technique, a local model is learned at each node
and then points which are outside this model are sent to the sink node
along with the model. These local outliers are then validated and the
global set is determined.
Two other approaches have been explored for classification in WSNs.
Bayesian approaches such as naive bayes, dynamic bayes and bayesian
belief propagation models have been used by Elnahrawy and Nath [23]