DOES DATASET SPLITTING IMPACT ARABIC TEXT CLASSIFICATION MORE THAN PREPROCESSING? AN EMPIRICAL ANALYSIS IN BIG DATA ANALYTICS

Text classification in Arabic faces numerous challenges because of the difficult structure of Arabic morphology and the absence of agreement about dataset preparation methods and splitting techniques. Research studies mostly examine feature engineering and preprocessing techniques yet fail to investigate properly the relationship between dataset splitting proportions and classification results. Research analyzes how the ratios between training and testing datasets (10%-90%, 20%-80% and 30%-70%) affect classification outcomes together with normalization, stop word filtering before stem, and tokenize operations. Numerous evaluation tests are conducted using 111,728 documents extracted from three Arabic newspapers that cover topics related to sports, politics and culture, economy, and diversity. The evaluation includes six machine-learning algorithms which are Random Forest, Support Vector Machine (SVM), Logistic Regression, Naïve Bayes, K-Neighbors, and Decision Tree. The analysis shows that dividing the dataset impacts model performance more extensively than applying preprocessing to datasets especially when operating on large test sets. The application of genetic preprocessing resulted in minor accuracy enhancements for the Decision Tree model at +2.30% while it produced precision gains of +2.48% but SVM achieved only +0.64% sensitivity improvement through normalization techniques. Stemming and tokenization delivered the best preprocessing results because Arabic possesses numerous morphological forms. The data shows trained accuracy depends on well-managed training/testing splits, which creates an important contradiction regarding traditional preprocessing methods. This analysis adds meaningful value to Arabic text classification research by demonstrating quantitative data about the effect of dataset division techniques through practical recommendations for enhancing classification performance in big data analytics. The research results possess relevant applications in media organizations because they help optimize Arabic text processing in education systems while also enhancing business intelligence operations.