An enhanced feature selection technique for classification of group based holy Quran verses

This thesis is about proposing an enhanced feature selection technique for text classification applications. Text classification problem is primarily applied in document labeling. However, the major setbacks with the existing feature selection techniques are high computational runtime associated with wrapper-based FS techniques and low classification accuracy performance associated with filter-based FS techniques. Therefore, in this study, a hybrid feature selection technique is proposed. The proposed FS technique is a combination of filter-based information gain (IG) and wrapper-based CFS algorithms. The purpose of combining these two FS algorithms is to achieve both high classification accuracy performance (wrapper) at lower computational runtime (filter). The study also developed a group-based Quran dataset to improve on the understanding and analysis of the textual data (Quranic verses). The group-based dataset is a combination of Holy Quran translation and commentary (tafsir). The Quranic verses were selected from two chapters, Surah Al- Baqarah and Surah Al-Anaam. The verses are classified into three categories: Faith, Worship, and Etiquette. In the experiment, six feature selection algorithms were applied: Information Gain (IG), Chi-square (CH), Pearson Correlation Coefficient (PCC), ReliefF, Correlation-based (CFS), and the proposed IG-CFS algorithms. The textual data (Quranic verses) were preprocessed using StringtoWordVector with weighted Term Frequency-Inverse Document Frequency (TF-IDF). Meanwhile, the classification phase has involved four algorithms: Naïve Bayes (NB), k-Nearest Neighbor (k-NN), Support Vector Machine (LibSVM), and Decision Trees (J48). The experiment results were evaluated based on two established performance metrics in text classification: Accuracy and Area under Receiver Operating Characteristics (ROC) curve (AUC). The proposed hybrid feature selection technique has shown promising results in terms of Accuracy and Area under Receiver Operating Characteristics (ROC) curve (AUC) by achieving at a lower computational runtime (3.89secs) Accuracy of 94.5% and AUC of 0.944 with the group-based Quran dataset