Comparative Evaluation of Machine Learning and Hyperparameter Optimization Methods for Low-Cost CO2 Sensor Calibration in Terms of Performance and Computational Cost

Low-cost sensors (LCSs) are increasingly used in air quality monitoring because of their affordability and scalability; however, their limited accuracy necessitates reliable calibration approaches. Although machine learning (ML)-based calibration methods have shown promising results, direct comparisons of hyperparameter optimization (HPO) strategies remain challenging due to differences in datasets, search spaces, and optimization budgets. In this study, ML models and HPO methods were evaluated within a standardized experimental framework developed on the AQ-MultiCal platform. Grid Search (GS), Random Search (RS), and Bayesian Optimization (BO) were implemented using identical hyperparameter search spaces and equal iteration budgets across both short-term and long-term real-world CO2 datasets obtained from five low-cost NDIR-based sensors. The results showed that tree-based models achieved strong baseline performance, whereas the k-nearest neighbors (kNN) model demonstrated the greatest improvement after optimization. The optimized kNN model reduced the average RMSE from 77.4 ppm to 54.4 ppm for the short-term dataset and from 37.3 ppm to 26.2 ppm for the long-term dataset. Although the HPO methods achieved comparable predictive accuracy, substantial differences were observed in computational cost. The proposed framework enables fair and reproducible comparison of HPO strategies while balancing predictive performance and computational efficiency in real-world sensor calibration applications.