Relief-Based Feature Selection for ANN-Driven Air Pollution Prediction and School Safety Policy in India

Abstract Public schools educate diverse socioeconomic groups, making air quality monitoring vital for students’ health and learning. These students are vulnerable to pollutants such as NO₂, SO₂, and PM₂.₅. Therefore air pollutants (AP) prediction helps in children’s health protection, learning enhancement, preventive measures, policy and awareness. Predictive models using ANNs, feature selection poses challenges due to data’s high dimensionality which increases over fitting risk and model complexity. Irrelevant features introduce noise, thereby reducing accuracy. So exploration of feature selection technique on ANN model accuracy remains an important area. This study explore relief algorithm (RA) for input selection to assess NO₂, SO₂, and PM₂.₅ prediction accuracy using multilayer feed forward neural network (MLFFNN), radial basis function neural network (RBFNN), generalized regression neural network (GRNN) and multi linear regression (MLR) model. RA ranks variables like wind direction, vertical wind speed, and solar radiation. It significantly enhanced model performance, increasing R 2 by 82.79% (SO₂) in RA-MLFFNN-1 and 77.37% (NO₂) in RA-MLFFNN-2. However, PM₂.₅ prediction using RA-MLFFNN-3 saw reduced accuracy, indicating model-specific effectiveness. The findings offer key insights for improving school air quality policy.