Distance-Based Meta-Features for Arabic Text Classification

Text classification (TC) is the process by which the computer has the ability to provide a label to a given text based on its content. Term Frequency–Inverse Document Frequency (TF-IDF) is one of the popular methods used for Arabic text representation. The high number of dimensions and sparseness are among the main issues faced by the TF-IDF method, requiring large space storage, high computational costs and the risk of overfitting. In this paper, we focus on four distance-based meta-features: CosKNN, L2KNN, CosCent and L2Cent derived from the TF-IDF representations, as a dimensionality reduction method for Arabic text classification. Four well-known classifiers are used in the present work: K-Nearest Neighbors, Logistic Regression, Support Vector Machines and Random Forest to evaluate the impact of these distance-based meta-features on the classification performance. The obtained results prove that the proposed dimensionality reduction method improves the classification accuracy in 50% of the cases and speed up the training phase (between 8x and 1764x faster) when compared to the original TF-IDF. As far as we know, distance-based meta-features are used for Arabic text classification for the first time.