Abstract
Feature vectors are extracted to represent objects in many classification tasks, such as text classification. Due to the high dimensionality of these raw feature vectors, the classification efficiency and accuracy are reduced. Therefore, reducing the size of feature vectors by selecting the relevant features that better represent the objects is an important aspect in text classification. Feature selection not only reduces the dimensionality of the feature vectors, but also produces more efficient classification models with higher predictive power. In this paper, we propose ARTC, which is an effective feature selection method that is based on the extraction of association rules to classify text documents. The extracted association rules discover the hidden relationships and correlations between the relevant words within the textual documents of a class and a cross different classes. Consequently, each class of documents is represented by a small set of contrasting features that are more effective in text classification. Our experiments show that ARTC outperforms other relevant techniques in terms of classification performance and efficiency.