Granular computing-based approach of rule learning for binary classification

Rule learning is one of the most popular types of machine-learning approaches, which typically follow two main strategies: ‘divide and conquer’ and ‘separate and conquer’. The former strategy is aimed at induction of rules in the form of a decision tree, whereas the latter one is aimed at direct induction of if–then rules. Due to the case that the divide and conquer strategy could result in the replicated sub-tree problem, which not only leads to overfitting but also increases the computational complexity in classifying unseen instances, researchers have thus been motivated to develop rule learning approaches through the separate and conquer strategy. In this paper, we focus on investigation of the Prism algorithm, since it is a representative one that follows the separate and conquer strategy, and is aimed at learning a set of rules for each class in the setting of granular computing, where each class (referred to as target class) is viewed as a granule. The Prism algorithm shows highly comparable performance to the most popular algorithms, such as ID3 and C4.5, which follow the divide and conquer strategy. However, due to the need to learn a rule set for each class, Prism usually produces very complex rule-based classifiers. In real applications, there are many problems that involve one target class only, so it is not necessary to learn a rule set for each class, i.e., only a set of rules for the target class needs to be learned and a default rule is used to indicate the case of non-target classes. To address the above issues of Prism, we propose a new version of the algorithm referred to as PrismSTC, where ‘STC’ stands for ‘single target class’. Our experimental results show that PrismSTC leads to production of simpler rule-based classifiers without loss of accuracy in comparison with Prism. PrismSTC also demonstrates sufficiently good performance comparing with C4.5.