Pruning methods for rule induction
Othman, O M (2016) 'Pruning methods for rule induction'.
Machine learning is a research area within computer science that is mainly concerned with discovering regularities in data. Rule induction is a powerful technique used in machine learning wherein the target concept is represented as a set of rules. The attraction of rule induction is that rules are more transparent and easier to understand compared to other induction methods (e.g., regression methods or neural network). Rule induction has been shown to outperform other learners on many problems. However, it is not suitable to handle exceptions and noisy data in training sets, which can be solved by pruning.
This thesis is concerned with investigating whether preceding rule induction with instance reduction techniques can help in reducing the complexity of rule sets by reducing the number of rules generated without adversely affecting the predictive accuracy.
An empirical study is undertaken to investigate the application of three different rule classifiers to datasets that were previously reduced with promising instance-reduction methods. Furthermore, we propose a new instance reduction method based on Ant Colony Optimization (ACO). We evaluate the effectiveness of this instance reduction method for k nearest neighbour algorithms in term of predictive accuracy and amount of reduction. Then we compared it with other instance reduction methods. We show that pruning classification rules with instance-reduction methods lead to a statistically significant decrease in the number of generated rules, without adversely affecting performance. On the other hand, applying instance-reduction methods enhances the predictive accuracy on some datasets. Moreover, the results provide evidence that: (1) our proposed instance reduction method based on ACO is competitive with the well-known k-NN algorithm; (2) the reduced sets computed by our method offers better classification accuracy than those obtained by the compared algorithms.