🚱 Water-Potability-Classification 🚰

Table of Content

1. Project Overview
2. Dataset Overview
3. Depedencies
4. Methodology
5. Result and Insight
6. Author

Project Overview

The Water Potability Classification project focuses on building a machine learning model to determine whether water is safe for consumption based on key physicochemical properties. The primary goal of this project is developing a reliable and efficient classification system that can assist in identifying potable water sources. By providing an automated solution, this project benefits consumers by enabling quicker and more accurate assessments of water quality, empowering communities, water management agencies, and environmental researchers to make informed decisions about water safety.

Dataset Overview

The Water Quality is a publicly accessible dataset provided by Aditya Kadiwal on Kaggle, designed to classify water potability based on various chemical and physical parameters.

Feature	Data Type	Description
ph	Float	pH value of water (range: 0 to 14).
Hardness	Float	Capacity of water to precipitate soap, measured in mg/L.
Solids	Float	Total dissolved solids in water, measured in ppm.
Chloramines	Float	Amount of Chloramines in water, measured in ppm.
Sulfate	Float	Amount of Sulfates dissolved in water, measured in mg/L.
Conductivity	Float	Electrical conductivity of water, measured in μS/cm.
Organic_carbon	Float	Amount of organic carbon in water, measured in ppm.
Trihalomethanes	Float	Amount of Trihalomethanes in water, measured in μg/L.
Turbidity	Float	Measure of the light-emitting property of water, measured in NTU.
Potability	Integer	Indicates if water is safe for human consumption: 1 = Potable, 0 = Not potable.

Depedencies

Below are the Python libraries used in this project, along with their functionalities:

• pandas: For data manipulation and analysis.
• matplotlib.pyplot: For creating static visualizations.
• seaborn: For creating informative and attractive statistical graphics.
• numpy: For numerical operations and handling arrays.
• scikit-learn:
  • train_test_split: To split data into training and testing sets.
  • RandomForestClassifier, DecisionTreeClassifier, GaussianNB, LogisticRegression: For implementing machine learning models.
  • accuracy_score, classification_report, confusion_matrix, roc_curve, auc, precision_recall_curve: For model evaluation and performance metrics.
  • SelectKBest, f_classif, chi2: For feature selection.
  • MinMaxScaler: For feature scaling.
  • GridSearchCV, RandomizedSearchCV: For hyperparameter tuning.
  • learning_curve: For plotting learning curves.
  • BaseEstimator, ClassifierMixin: For creating custom classifiers.
• imblearn:
  • SMOTE: For handling imbalanced datasets via oversampling.
• xgboost:
  • XGBClassifier: For implementing gradient boosting algorithms.
• mlxtend:
  • SequentialFeatureSelector (SFS): For sequential feature selection.
• sklearn.feature_selection:
  • RFE: For recursive feature elimination.
• lime:
  • lime_tabular: For generating interpretable explanations of machine learning model predictions using the LIME (Local Interpretable Model-Agnostic Explanations) framework.

These libraries collectively enable data preprocessing, model building, evaluation, optimization, and interpretability for the water potability classification project.

Methodology

The methodology for the Water Potability Classification project is divided into the following steps:

1. Library Import
   Necessary Python libraries are imported to facilitate data manipulation, visualization, preprocessing, modeling, evaluation, and interpretability.

2. Exploratory Data Analysis (EDA)
   Initial analysis is conducted to understand the data distribution, identify missing values, and explore relationships between features and the target variable (Potability). Visualizations like histograms, boxplots, and correlation heatmaps are used for insights.

3. Preprocessing and Transformation

   • Data Cleaning: Handle missing values, outliers, and inconsistencies in the dataset.
   • Data Transformation: Normalize numerical features to ensure uniform scaling.
   • Feature Selection: Apply statistical methods such as Chi-square (Chi2), Sequential Feature Selector (SFS), and Recursive Feature Elimination (RFE) to identify important features.
   • Resampling: Use SMOTE (Synthetic Minority Oversampling Technique) to address class imbalance.
   • Feature Selection on Resampled Data: Reapply feature selection techniques after balancing the dataset.

4. Training
   Train models using five different machine learning algorithms, including Random Forest, Logistic Regression, XGBoost, Decision Tree, and Naive Bayes. For each algorithm, experiments are conducted using:

   • Original data.
   • Resampled data.
   • Different feature selection techniques to compare performance.

5. Evaluation of the Best Model
   Evaluate model performance using metrics such as accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix. The best model is selected based on its performance across these metrics.

6. Model Interpretation with LIME
   Use the LIME (Local Interpretable Model-Agnostic Explanations) framework to interpret and explain the predictions of the best-performing model. LIME provides feature importance for individual predictions, enhancing transparency and trust in the model.

7. Error Analysis
   Analyze misclassified samples to understand patterns in prediction errors, identify potential limitations, and recommend improvements.

8. User Scenarios
   Provide use cases and scenarios where the model can be deployed, such as:

   • Monitoring water quality in real-time systems.
   • Assisting water management agencies in decision-making.
   • Providing rapid water safety assessments for communities and researchers.

This structured approach ensures the project achieves its goals of creating an accurate, interpretable, and user-oriented water potability classification system.

Result and Insight

1. Best Algorithm
   The best-performing algorithm in this study is the Random Forest, which achieved the highest performance among the tested algorithms. It demonstrated the best performence for both the original and resampled datasets.

2. Best Features
   The top six features contributing most to the model's performance, as identified through BFE's feature selection techniques, are:

   • Hardness
   • Solids
   • Chloramines
   • ph
   • Sulfate
   • Conductivity

3. Performance Metrics
   The Random Forest model achieved the following metrics on the test set:

   • Accuracy: 74%
   • Precision: 72%
   • Recall: 65%
   • F1-Score: 74%

4. Insights from Feature Importance

   • Hardness and Solids: Highly correlated with water's mineral content, which plays a crucial role in determining its usability.
   • Chloramines: A key indicator of water treatment processes, directly affecting potability.
   • ph: Strongly linked to acidity or alkalinity, a primary factor in water safety.
   • Sulfate and Conductivity: Indicators of chemical composition and electrical properties, influencing water purity.

5. Interpretation with LIME
   LIME interpretations showed that specific feature values (e.g., high chloramines levels or extreme pH values) consistently influenced the model's classification, providing valuable insights for practical applications like real-time water quality monitoring.

6. Error Analysis
   Misclassifications mostly occurred in borderline cases where feature values were near threshold levels (e.g., marginal pH or sulfate levels). This suggests that additional data or refined thresholds could further improve accuracy.

7. User Scenarios and Deployment
   The model is suitable for deployment in:

   • Automated water quality monitoring systems.
   • Mobile applications for community-based water testing.
   • Research projects analyzing regional water quality trends.

Author

(Muhammad Hadi Nur Fakhri)[https://github.com/NurFakhri)