EARM: A MULTI-LEVEL ENSEMBLE APPROACH TO ASSOCIATION RULE MINING FOR ENHANCED TEXT CLASSIFICATION

Introduction: Text classification is a high-dimensionality problem; therefore, feature selection is still one of the crucial problems in this domain. Association rule mining has proven to be a useful way to identify important features by understanding how they work together, but current methods do not fully take advantage of its capabilities. Method: In this paper, we propose an Enhanced Association Rule Mining (EARM) framework for feature selection on text classification tasks. EARM's process for finding rules works at different levels, helps distinguish between specific and overall rules, enhances traditional association rule mining, uses a new method to balance support, confidence, and lift metrics, and combines different rule mining techniques to make it more reliable. We tested EARM on standard datasets, and the findings show that it works better than the best current feature selection methods, including a new method for frequent correlated items. Our approach demonstrates a 3.7% accuracy increase in classification with an additional 94% reduction in dimensionality compared to the complete feature set. Conclusions: These powerful association rule-mining primitives imply that these advanced, improved techniques are usable as feature selectors in text classification.