AquaAugmentor: A Novel Feature Augmentation Algorithm for Water Potability Prediction

Technical Summary: AquaAugmentor for Water Potability Prediction

Problem Statement
Access to safe drinking water is a critical global challenge, with over two billion people living in water-stressed regions. Traditional water quality testing is often time-consuming and requires specialized equipment. While machine learning (ML) and deep learning (DL) offer rapid, cost-effective alternatives for classifying water potability based on chemical and physical parameters, these models face significant hurdles when applied to low-dimensional datasets. Existing approaches often struggle with the complexity and variability of water source data, leading to inconsistent predictive performance. Furthermore, many studies report high accuracies that are difficult to reproduce in real-world scenarios due to data noise and fluctuation.

Methodology
This study proposes AquaAugmentor, a novel feature augmentation algorithm designed to enhance the predictive performance of ML and DL models on low-dimensional water quality datasets. The methodology proceeds through the following stages:

1. Dataset Preparation: The study utilizes a dataset of 3,276 samples containing nine chemical attributes: pH, Hardness, Solids, Chloramines, Sulfate, Conductivity, Organic Carbon, Trihalomethanes, and Turbidity. The data includes 1,278 potable and 1,998 non-potable samples. Missing values were addressed via mean imputation.
2. Feature Augmentation (AquaAugmentor): The algorithm expands the original 9-feature set into a 62-feature set through three specific mechanisms:
   • Polynomial Features: Generating combinations of existing features up to a specified degree.
   • Statistical Features: Calculating sample-wise mean, standard deviation, minimum, and maximum values.
   • Domain-Specific Features: Creating new features based on domain knowledge, specifically ratios between specific pairs of original chemical attributes (e.g., pH to Sulfate ratios).
3. Data Processing: The augmented dataset undergoes class balancing using the Synthetic Minority Over-sampling Technique (SMOTE) to address the imbalance between potable and non-potable classes. The data is then split into training (70%), validation (15%), and test (15%) sets, followed by normalization using StandardScaler.
4. Model Evaluation: A comprehensive suite of models was evaluated, including traditional ML algorithms (Linear Regression, Random Forest, SVM, KNN, Naive Bayes, Logistic Regression, AdaBoost, XGBoost, GBM, CatBoost) and deep learning architectures (1-D CNN, LSTM, Autoencoder, Variational Autoencoder, TabNet, TabTransformer, and a hybrid CNN-LSTM). Each model was tested with and without the AquaAugmentor algorithm, and performance was measured using Test Accuracy and Area Under the Curve (AUC) scores.

Key Results
The application of AquaAugmentor resulted in consistent performance improvements across the majority of tested models:

• Random Forest: Accuracy increased from 65.71% to 72.62%, and AUC improved from 0.68 to 0.80.
• XGBoost: Accuracy rose significantly from 54.27% to 71.83%, with AUC improving from 0.55 to 0.80.
• Deep Learning Models: The LSTM model saw accuracy improve from 57.93% to 66.67% (AUC: 0.53 to 0.73), while the Autoencoder improved from 57.93% to 67.67% (AUC: 0.51 to 0.76).
• Statistical Validation: T-tests confirmed that the disparities in the augmented features between potable and non-potable samples were statistically significant (p < 0.05), validating the efficacy of the feature engineering.

The authors note that while some literature claims accuracies exceeding 90%, such results are often not reproducible in practical, noisy environmental contexts. In contrast, the improvements achieved by AquaAugmentor are practical and tangible, offering robust enhancements to model reliability without claiming unrealistic perfection.

Significance and Claims
The paper positions AquaAugmentor as a valuable framework for improving water quality assessment within the context of Industry 5.0 and sustainable development. Its primary significance lies in:

• Enhancing Low-Dimensional Data: It provides a robust solution for improving predictive performance specifically on datasets with limited feature counts, a common constraint in environmental monitoring.
• Practical Applicability: Unlike approaches that may overfit to specific datasets, this method offers a generalized strategy to enrich feature spaces, making ML and DL models more suitable for real-world deployment where data is often noisy.
• Decision Support: The findings aim to assist researchers, policymakers, and public health officials in making informed decisions regarding water safety by providing more reliable classification tools.

The authors conclude that while the algorithm introduces increased computational complexity and training time, the trade-off yields models with greater robustness and reliability for ensuring safe drinking water access. Future work suggested by the authors includes investigating adaptive feature augmentation to dynamically adjust complexity and integrating the algorithm with real-time IoT sensor networks.