A new classification model with a multi-layer dimensionality reduction approach

Documents categorization has evolved into a pertinent study topic, adding new challenges to data mining and information retrieval fields. However, classification algorithms suffer from the large sparse matrix generated from the processed dataset. Therefore, reducing data dimensionality is considered an essential phase in the whole process of classification. This phase aims to remove irrelevant features while enhancing the classifier’s performance. Several feature extraction and selection methods have been introduced and discussed. Still, applying these methods in Arabic topic categorization is not well-studied and left behind in comparison with various languages like English. In this paper, we highlight the benefit of using a hybrid multi-layer dimension reduction approach, especially for Arabic topic classification. The procedure proposed includes four layers of feature extraction and selection. The first two layers apply linguistic methods, while the last two layers use statistical analyses. Then, a bagging classifier model is implemented using the Support Vector Machines (SVM) as its base model. Finally, we evaluated the classifier performance using several measurable criteria such as precision, recall, accuracy, and F-score. According to the literature, the presented model reached an average accuracy that is among the best results reported in previous studies.