Feature Selection and the Preservation of Infrequent and Highly Significant Attributes in the Context of Arabic Text Mining

Effective feature selection is a key component for building an efficient automatic document classifier. We regularly encounter in the Arabic literatureespecially the scientific oneinfrequent non-Arabic words that are eliminated by practice during the pre-processing phase. Although infrequent, those words are highly pertinent to their documents and, thus, can contribute to build a more efficient classification model and enforce the subjectivity of the decision taken by the classifier. Therefore, we propose in this paper four different feature selection solutions that allow both preserving a maximum number of those words and getting satisfactory classification accuracy.