<![CDATA[Purpose – This study aims to examine whether self-reported attention-related difficulty categories among social media users can be classified using a leakage-free machine learning framework. It addresses the risk of inflated performance in survey-based classification by excluding the same items used to construct the target label from the predictor set.Methods – The study used the public Social Media and Mental Health (SMMH) Kaggle dataset with 478 valid respondents. A three-class label was constructed from Q10, Q12, and Q14 using percentile thresholds (P33 = 9.0; P66 = 12.0), producing High (n = 208), Medium (n = 152), and Low (n = 118) categories. These label-generating items were excluded from predictors. The remaining variables were processed in a scikit-learn Pipeline using MinMax scaling, ordinal encoding, and One-Hot Encoding. Multi-kernel SVM models and five baseline classifiers were evaluated using a stratified 70:30 split, cross-validation, F1 metrics, balanced accuracy, and permutation importance.Findings – Random Forest achieved the highest performance, with 63.19% accuracy and 62.26% weighted F1. Linear SVM was the best SVM model, achieving 61.81% accuracy, 60.08% weighted F1, 58.99% macro F1, and 59.11% balanced accuracy. The strongest predictors were Restless Without Social Media, Use Without Purpose, and Interest Fluctuation.Research implications – The findings are preliminary, dataset-specific, and based on a survey-derived composite label whose internal reliability still requires validation.Originality – This study contributes a leakage-controlled classification approach for analyzing attention-related survey categories.]]>
