Robust Feature Management for Gaming Disorder Classification Using Modified Partial Instance Reduction and Fine-Tune Attribute-Weighted Naïve Bayes

Abstract

The dynamic nature of the social environment can influence symptom patterns in mental illnesses, thereby affecting their classification. Gaming disorder (GD), recently recognized in ICD-11 by the World Health Organization, is currently identified based on 44 symptoms derived from expert assessments. However, these symptoms may shift as social behaviors evolve, necessitating classification models that remain robust despite changes in dataset attributes. This study proposes a robust model—fine-tuned attribute-weighted Naïve Bayes (FTAWNB) with modified partial instance reduction (mPIR)—to address this issue. Two test scenarios were conducted: the addition of one new attribute and the removal of four attributes from the GD dataset, applied to both the original and updated datasets. The results indicate that FTAWNB with modified PIR enhances classification performance. On the original dataset, accuracy increased by 1.28% (Scenario 1) and 1.4% (Scenario 2). On the updated dataset, the model maintained 99.74% accuracy (Scenario 1) and improved by 1.4% (Scenario 2). These findings demonstrate that the integration of modified PIR improves model stability in dynamic feature environments, thereby contributing to more reliable mental health classification systems.