Feature Selection Methods Using RBFNN Based on Enhance Air Quality Prediction: Insights from Shah Alam

• Published in: INTERNATIONAL JOURNAL OF ADVANCED COMPUTER SCIENCE AND APPLICATIONS
• Main Authors: Arafin, Siti Khadijah; Ul-Saufie, Ahmad Zia; Ghani, Nor Azura Md; Ibrahim, Nurain
• Format: Article
• Language: English
• Published: SCIENCE & INFORMATION SAI ORGANIZATION LTD 2024
• Subjects: Computer Science
• Online Access: https://www-webofscience-com.uitm.idm.oclc.org/wos/woscc/full-recordWOS:001378221800001
• ISSN: 2158-107X; 2156-5570

This study examines the predictive efficiency of several feature selection approaches in air quality models aimed to predict next-day PM2.5 concentrations in Shah Alam, Malaysia. Air pollution in urban areas is a significant public health concern, and accurate prediction models are essential for timely interventions. However, determining the most important parameters to include in these models remains difficult, especially in complex urban areas with several pollution sources. To address this, we employed three different feature selection methods and applied them to a dataset comprising 43,824 air quality data points provided by the Department of Environmental Malaysia. The data set contained ten variables, such as gas pollutants and meteorological indicators. Each feature selection approach determined top eight variables to include in a Radial Basis Function Neural Network (RBFNN) model. The results showed that ReliefF outperformed Lasso and mRMR in terms of accuracy, specificity, precision, F1 Score, and AUROC, making it the most effective feature selection method for this study. This study contributes to the body of knowledge on air quality modelling by emphasising the relevance of using proper feature selection techniques that are suited to the specific characteristics of the dataset and urban area. Furthermore, it proposes that future study should look into the use of Relief-FRBFNN in other settings, such as suburban and rural areas, as well as hybrid feature selection approaches to improve prediction performance across several context.