An adaptive ant colony optimization algorithm for rule-based classification

• Main Author: Al-Behadili, Hayder Naser Khraibet
• Format: Thesis
• Language: English; English; English
• Published: 2020
• Subjects: QA Mathematics
• Online Access: https://etd.uum.edu.my/8786/1/Deposit%20Permission_s901983.pdf; https://etd.uum.edu.my/8786/2/s901983_01.pdf; https://etd.uum.edu.my/8786/3/s901983_references.docx; https://etd.uum.edu.my/8786/

Classification is an important data mining task with different applications in many fields. Various classification algorithms have been developed to produce classification models with high accuracy. Differing from other complex and difficult classification models, rules-based classification algorithms produce models which are understandable for users. Ant-Miner is a variant of ant colony optimisation and a prominent intelligent algorithm widely use in rules-based classification. However, the Ant-Miner has overfitting and easily falls into local optima problems which resulted in low classification accuracy and complex classification rules. In this study, a new Ant-Miner classifier is developed, named Adaptive Genetic Iterated-AntMiner (AGI-AntMiner) that aims to avoid local optima and overfitting problems. The components of AGI-AntMiner includes: i) an Adaptive AntMiner which is a prepruning technique to dynamically select the appropriate threshold based on the quality of the rules; ii) Genetic AntMiner that improves the post-pruning by adding/removing terms in a dual manner; and, iii) an Iterated Local Search-AntMiner that improves exploitation based on multiple-neighbourhood structure. The proposed AGI-AntMiner algorithm is evaluated on 16 benchmark datasets of medical, financial, gaming and social domains obtained from the University California Irvine repository. The algorithm’s performance was compared with other variants of Ant-Miner and state-of-the-art rules-based classification algorithms based on classification accuracy and model complexity. Experimental results proved that the proposed AGI-AntMiner algorithm is superior in two (2) aspects. Hybridization of local search in AGI-AntMiner has improved the exploitation mechanism which leads to the discovery of more accurate classification rules. The new pre-pruning and postpruning techniques have improved the pruning ability to produce shorter classification rules which are easier to interpret by the users. Thus, the proposed AGI-AntMiner algorithm is capable in conducting an efficient search in finding the best classification rules that balance the classification accuracy and model complexity to overcome overfitting and local optima problems.