Abstract:
The primary goal of this research is to investigate the suitability of ant colony
optimization, a swarm intelligence based meta-heuristic developed by mimicking some
aspects of the food foraging behavior of ants, for building accurate and comprehensible
classifiers which can be learned in reasonable time even for large datasets. Towards this
end, a novel classification rule discovery algorithm called AntMiner-C and its variants
are proposed. Various aspects and parameters of the proposed algorithms are investigated
by experimentation on a number of benchmark datasets. Experimental results indicate
that the proposed approach builds more accurate models when compared with commonly
used classification algorithms. It is also computationally less expensive than previously
available ant colony algorithm based classification rules discovery algorithms.
A hybrid classifier using ant colony optimization is also proposed that combines
association rules mining and supervised classification. Experiments show that the
proposed algorithm has the ability to discover high quality rules. Furthermore, it has the
advantage that association rules of each class can be mined in parallel if distributed
processing is used. Experimental results demonstrate that the proposed hybrid classifier
achieves higher accuracy rates when compared with other commonly used classification
algorithms.
A feature subset selection algorithm is also proposed which is based on ant colony
optimization and decision trees. Experiments show that better accuracy is achieved if the
subset of features selected by the proposed approach is used instead of full feature set and
number of rules is also decreased substantially.
