Machine learning methods for imbalanced data set for prediction of faecal contamination in beach waters

Area of Science:

• Environmental Science
• Public Health
• Data Science

Background:

• Statistical models aid in managing health risks at recreational beaches.
• Predicting extreme water contamination events is difficult due to imbalanced datasets where rare events are infrequent.

Purpose of the Study:

• To evaluate machine learning techniques and metrics for modeling imbalanced data in predicting water contamination.
• To identify optimal methods for anticipating water contamination and mitigating health risks.

Main Methods:

• Utilized simulated datasets and a real-world database of fecal coliform abundance from 21 Uruguayan beaches over 10 years.
• Tested various machine learning algorithms and data pre-treatment methods like upsampling and SMOTE.
• Assessed model performance using True Positive Rate (TPR) and False Positive Rate (FPR), not just accuracy.

Main Results:

• Most machine learning techniques require data pre-treatment (e.g., upsampling) to enhance performance on imbalanced datasets.
• Stratified Random Forest demonstrated superior performance, improving TPR by 50% compared to baseline.
• Support Vector Machines with SMOTE and Adaboost with SMOTE also yielded strong results.

Conclusions:

• Machine learning models are sensitive to data imbalance, necessitating specific pre-processing techniques.
• TPR and FPR are more reliable metrics than accuracy for evaluating models with imbalanced data.
• Combining modeling strategies is crucial for improving the prediction of water contamination events and safeguarding public health.