A Hybrid Machine Learning Framework for Water Quality Index Prediction Using Feature-Based Neural Network Initialization

This project develops a hybrid machine learning framework for water quality index prediction, utilizing a dataset of 3,276 samples with nine physicochemical parameters to classify water as potable or non-potable. After addressing class imbalance with SMOTE (Synthetic Minority Over-sampling Technique) and standardizing features, the framework implements two innovative architectures: the Adaptive Attention-Guided Hybrid Ensemble (AAGHE) , which combines an autoencoder with CatBoost (Categorical Boosting), ExtraTrees (Extremely Randomized Trees), and HistGradientBoosting (Histogram-based Gradient Boosting) classifiers alongside an attention-based meta-learner to achieve 99.33% accuracy; and the Dual-Transformer Cross-Fusion Network (DTCF-Net) , which uses parallel transformer blocks with cross-attention fusion to capture complex feature interactions, achieving 68.81% accuracy. The system includes a deployment-ready inference pipeline that generates actionable treatment suggestions based on parameter threshold violations, supporting water safety management, public health monitoring, and timely intervention through rigorous evaluation metrics.