University Of Tehran Press Pollution 2383-451X 11 4 2025 10 01 Comparative Analysis of Machine Learning Models Vis-a-Vis Regularization Models on Predictive Ability of Pollution Levels in Bangalore 1419 1432 103812 10.22059/poll.2025.393694.2892 EN Vivekanand Venkataraman Department of Mechanical Engineering, BMS Institute of Technology & Management, Bengaluru – 54, Affiliated to Visvesvaraya Technological University Belagavi -18, Karnataka, India Mohan Babu G.N AMC Engineering College, Bengaluru – 560083, Affiliated to Visvesvaraya Technological University Belagavi -18, Karnataka, India Sathish Kumar K.M Department of Mechanical Engineering, BMS Institute of Technology & Management, Bengaluru – 54, Affiliated to Visvesvaraya Technological University Belagavi -18, Karnataka, India Journal Article 2025 04 18 The work brings in various dimensions to understand the importance of machine learning models in terms of predicting and understanding the variables which affect the concentration of Particulate matter PM2.5 and PM10 for Bangalore city. In this work Metrological variables, Pollutants are considered as inputs. In addition, the work highlights the differences achieved in terms of performance metric especially in terms of error variance and prediction power for particulate matter using Regularization, Bagging and Boosting techniques. It specifically brings about areas where these techniques can perform well and underperform. The work also compares how the models performed with and without features such as seasons. It was noticed the order of feature importance differed for regularization, boosting and bagging models. It was noticed that  Boosting techniques such as Xgboost had lower RMSE(9.1), MAPE (15.75) and higher R2 values (.72) for PM2.5 than other models however overfitting was noticed. However random forest had a lower R2 (.64) compared to boosting and RMSE (10.35) and MAPE (22.45) were slightly larger and tendency to overfit was lower.  To understand further, a new approach was created to diagnose where exactly these models perform well and underperform. The Absolute values were divided into percentile values and correlation was investigated with respect to error or residual values, adding to the uniqueness of this work. It was found that Extreme values tend to be correlated to larger residual values and those within the normal percentile range have lesser residual values.   https://jpoll.ut.ac.ir/article_103812_d386e3414aa3c26520375ae7c29c6bf7.pdf