Key insights into recommended SMS spam detection datasets.

Area of Science:

• Computer Science
• Information Security

Background:

• Short Message Service (SMS) spam presents significant security risks, including financial fraud and phishing.
• Existing SMS spam detection models often rely on various datasets, but their effectiveness and impact are under-evaluated.

Purpose of the Study:

• To assess the performance of ten SMS spam detection datasets.
• To evaluate datasets based on accuracy and qualitative factors for improved SMS spam detection models.

Main Methods:

• Utilized Decision Tree and Multinomial Naïve Bayes models to evaluate ten SMS spam detection datasets.
• Assessed datasets on accuracy, authenticity, class imbalance, feature diversity, metadata, and preprocessing needs.
• Conducted experiments with English and non-English stopword removal groups due to dataset multilingualism.

Main Results:

• Dataset 5 achieved a high qualitative score (3.8/5.0) due to feature diversity, real-world complexity, and balanced classes.
• Multinomial Naïve Bayes showed a detection rate of 86.10% on Dataset 5, indicating a challenging yet informative dataset.
• Qualitative assessment proved crucial in dataset evaluation beyond mere accuracy metrics.

Conclusions:

• Dataset 5 is recommended for future SMS spam detection research to foster more robust and adaptable models.
• Selecting datasets with high qualitative scores enhances research quality, model generalizability, and mitigates bias.
• Challenging datasets encourage the development of spam detection systems capable of handling diverse noise and ambiguity.