A Hybrid Filter based Multi-Class Classification Model for Large Spatial Databases

Abstract

As the size of the cyclone data increases, the essential cyclone patterns from large databases
become difficult to find and extract. The cyclone dataset includes a large number of high
dimensional space cyclone instances. In order to forecast cyclones on limited datasets, a large
number of cyclone prediction methods have been implemented in the literature. To classify
and predict the cyclone class using the limited cyclone data size, traditional cyclone
severity detection methods such as SVM, Naive bayes, regression models, neural network,
etc. are used. Such models are heavily dependent on cyclone characteristics and their
class distribution. These models are not effective in classifying the outliers using the patterns
of the cyclone and its places. These models also require high computational time and memory
for cyclone severity prediction as the size of the instances increases. In this approach, in
order to pre- process the cyclone data and to predict the cyclone location using the cyclone
severity, the cyclone region prediction and outlier detection (CRPCOD) mechanism using
the hybrid ensemble learning model is adopted. The PCA algorithm in the CRPCOD model
is used to filter the dimensions of the model for ensemble learning. Various base classifiers,
such as decision tree, SVM and neural network for cyclone location prediction, are enhanced
in the proposed ensemble learning model. Finally, using the outlier cyclone patterns for
decision making, cyclone change points are discovered. Experimental results showed that,
compared to the traditional cyclone severity prediction models, the proposed model has
high classification accuracy and low error rate.