A modified multi-class association rule for text mining

• Auteur principal: Al-Refai, Mohammad Hayel Abdel Karim
• Format: Thèse
• Langue: anglais; anglais
• Publié: 2015
• Sujets: QA71-90 Instruments and machines
• Accès en ligne: https://etd.uum.edu.my/5767/1/depositpermission_s91487.pdf; https://etd.uum.edu.my/5767/2/s91487_01.pdf

Classification and association rule mining are significant tasks in data mining. Integrating association rule discovery and classification in data mining brings us an approach known as the associative classification. One common shortcoming of existing Association Classifiers is the huge number of rules produced in order to obtain high classification accuracy. This study proposes s a Modified Multi-class Association Rule Mining (mMCAR) that consists of three procedures; rule discovery, rule pruning and group-based class assignment. The rule discovery and rule pruning procedures are designed to reduce the number of classification rules. On the other hand, the group-based class assignment procedure contributes in improving the classification accuracy. Experiments on the structured and unstructured text datasets obtained from the UCI and Reuters repositories are performed in order to evaluate the proposed Association Classifier. The proposed mMCAR classifier is benchmarked against the traditional classifiers and existing Association Classifiers. Experimental results indicate that the proposed Association Classifier, mMCAR, produced high accuracy with a smaller number of classification rules. For the structured dataset, the mMCAR produces an average of 84.24% accuracy as compared to MCAR that obtains 84.23%. Even though the classification accuracy difference is small, the proposed mMCAR uses only 50 rules for the classification while its benchmark method involves 60 rules. On the other hand, mMCAR is at par with MCAR when unstructured dataset is utilized. Both classifiers produce 89% accuracy but mMCAR uses less number of rules for the classification. This study contributes to the text mining domain as automatic classification of huge and widely distributed textual data could facilitate the text representation and retrieval processes.