Discovering Regression and Classification Rules with Monotonic Constraints Using Ant Colony Optimization

Data mining is a broad area that encompasses many different tasks from the supervised classification and regression tasks to unsupervised association rule mining and clustering. A first research thread in this thesis is the introduction of new Ant Colony Optimization (ACO)-based algorithms that tackle the regression task in data mining, exploring three different learning strategies: Iterative Rule Learning, Pittsburgh and Michigan strategies. The Iterative Rule Learning strategy constructs one rule at a time, where the best rule created by the ant colony is added to the rule list at each iteration, until a complete rule list is created. In the Michigan strategy, each ant constructs a single rule and from this collection of rules a niching algorithm combines the rules to create the final rule list. Finally, in the Pittsburgh strategy each ant constructs an entire rule list at each iteration, with the best list constructed by an ant in any iteration representing the final model. The most successful Pittsburgh-based Ant-Miner-Reg_PB algorithm, among the three variants, has been shown to be competitive against a well-known regression rule induction algorithm from the literature. The second research thread pursued involved incorporating existing domain knowledge to guide the construction of models as it is rare to find new domains that nothing is known about. One type of domain knowledge that occurs frequently in real world data-sets is monotonic constraints which capture increasing or decreasing trends within the data. In this thesis, monotonic constraints have been introduced into ACO-based rule induction algorithms for both classification and regression tasks. The enforcement of monotonic constraints has been implemented as a two step process. The first is a soft constraint preference in the model construction phase. This is followed by a hard constraint post-processing pruning suite to ensure the production of monotonic models. The new algorithms presented here have been shown to maintain and improve their predictive power when compared to non-monotonic rule induction algorithms.