Comparative Evaluation of Machine Learning and Hyperparameter Optimization Methods for Low-Cost CO₂ Sensor...

Area of Science:

• Environmental Science
• Sensor Technology
• Data Science

Background:

• Low-cost sensors (LCSs) are vital for scalable air quality monitoring but require accurate calibration.
• Machine learning (ML) shows promise for LCS calibration, yet comparing hyperparameter optimization (HPO) strategies is difficult due to varied experimental setups.

Purpose of the Study:

• To standardize the evaluation of ML models and HPO strategies for LCS calibration.
• To compare Grid Search, Random Search, and Bayesian Optimization using a unified framework.

Main Methods:

• Utilized the AQ-MultiCal platform for a standardized experimental framework.
• Implemented Grid Search, Random Search, and Bayesian Optimization on CO2 datasets from NDIR-based LCSs.
• Evaluated performance using identical hyperparameter search spaces and iteration budgets.

Main Results:

• Tree-based models provided strong baseline performance; k-nearest neighbors (kNN) showed the most significant improvement post-optimization.
• Optimized kNN reduced average RMSE from 77.4 ppm to 54.4 ppm (short-term) and 37.3 ppm to 26.2 ppm (long-term).
• HPO methods yielded similar predictive accuracy but varied considerably in computational cost.

Conclusions:

• The standardized framework allows fair comparison of HPO strategies for LCS calibration.
• Balancing predictive performance and computational efficiency is key for real-world sensor calibration.
• This study provides a reproducible method for selecting optimal HPO strategies for air quality monitoring.