💧 Water Potability Prediction

A machine learning pipeline that predicts whether water is safe to drink using physicochemical features. Four classifiers are trained, compared, and explained with SHAP. Deployed as an interactive web app via Streamlit.

---

📁 Project Structure

|-- water_quality.ipynb      # Main notebook (EDA → preprocessing → training → evaluation → SHAP)
|-- water_potability.csv     # Dataset (Kaggle — Water Potability, 3,276 samples)
|-- app.py                   # Streamlit deployment app (AquaSafe)
|-- presentation.pdf         # Presentation pdf file
|-- requirements.txt         # Python dependencies
|-- rf_model.pkl             # Saved Random Forest model
|-- scaler.pkl               # Saved RobustScaler

---

📊 Dataset

Source: Kaggle — Water Potability Dataset
File: water_potability.csv
Rows: 3,276  |  Features: 9 physicochemical properties + 1 binary target

Feature	Description	Unit
ph	Acidity/alkalinity of water	0 – 14
Hardness	Capacity to precipitate soap	mg/L
Solids	Total dissolved solids	ppm
Chloramines	Chloramines concentration	ppm
Sulfate	Sulfate concentration	mg/L
Conductivity	Electrical conductivity	μS/cm
Organic_carbon	Total organic carbon	ppm
Trihalomethanes	Trihalomethanes concentration	μg/L
Turbidity	Cloudiness of water	NTU
Potability	Target — 1 = potable, 0 = not potable	—

Class distribution: 1,998 not potable (61%) / 1,278 potable (39%) — imbalanced.
Missing values: ph (491), Sulfate (781), Trihalomethanes (162).

---

🔧 Pipeline

1. EDA

• Distribution histograms for all 9 features
• Correlation heatmap
• Class imbalance check

2. Preprocessing

• Imputation: KNN Imputer (k=5) for missing values
• Scaling: RobustScaler (robust to outliers)
• Split: 80/20 train/test, stratified by target
• SMOTE: Applied to training set only → balances to 1,598 samples per class

3. Models & Hyperparameter Tuning

GridSearchCV with 5-fold cross-validation, optimising F1 score.

Model	Hyperparameters Searched
Logistic Regression	solver=liblinear
Random Forest	n_estimators ∈ {100, 200}, max_depth ∈ {None, 10, 20}
Gradient Boosting	Default + random state
SVM (RBF)	C ∈ {0.1, 1, 10}, gamma ∈ {scale, auto}

---

📈 Results

Model	Accuracy	ROC-AUC	F1 (Potable)
Random Forest ✅	65%	0.6698	0.51
Gradient Boosting	58%	0.6526	0.54
SVM	59%	0.6334	0.49
Logistic Regression	~57%	~0.61	~0.47

Best model: Random Forest (highest ROC-AUC — selected for deployment and SHAP analysis).

---

🔍 SHAP Interpretability

SHAP TreeExplainer applied to the best Random Forest model reveals which features drive predictions most.

Key findings from the beeswarm plot:

• pH is the most influential feature — high pH pushes toward potable, low pH toward not potable
• Sulfate is the second most important — high values have inconsistent effects
• Hardness & Chloramines show moderate influence
• Turbidity, Trihalomethanes, Conductivity contribute minimally to predictions

Plots generated:

• Beeswarm plot — per-sample feature impact
• Bar chart — mean absolute SHAP values (global importance ranking)
• Waterfall plot — single-sample prediction explanation

---

🚀 Usage

Run the notebook

jupyter notebook water_quality.ipynb

Run the Streamlit app locally

pip install -r requirements.txt
streamlit run app.py

Live demo

👉 waterqualityclassification-project1.streamlit.app

---

⚠️ Notes & Limitations

• SMOTE is applied only on the training set to prevent data leakage into evaluation.
• Model performance is moderate (~0.67 AUC); water potability is inherently noisy with physicochemical features alone.
• The app uses WHO guideline thresholds for contextual warnings — these are advisory, not hard classification rules.
• Not intended for production use — always verify with certified laboratory testing before making public health decisions.