Classification by Frequent Association Rules

Abstract

Over the last two decades, Associative Classifiers have shown competitive performance in the task of predicting class labels. Along with the performance in accuracy, associative classifiers produce human-readable predictive rules which is very helpful to understand the decision process of the classifiers. Associative classifiers from early days suffer from the limitation requiring proper threshold value setting which is dataset-specific. Recently some studies eliminated that limitation by producing statistically significant rules. Though recent models showed very competitive performance with state-of-the-art classifiers, their performance is still impacted if the feature vector of the training data is very large. An ensemble model can solve this issue by training each base learner with a subset of the feature vector. In this study, we propose an ensemble model Classification by Frequent Association Rules (CFAR) using associative classifiers as base learners. In our approach, instead of using a classical ensemble and a voting method, we rank the generated rules based on predominance among base learners and select a subset of the rules for predicting class labels. We use 10 datasets from the UCI repository to evaluate the performance of the proposed model. Our ensemble approach CFAR eliminates the limitation of high memory requirement and runtime of recent associative classifiers if training datasets have large feature vectors. Among the datasets we used, along with increasing accuracy in most cases, CFAR removes the noisy rules which enhances the interpretability of the model.