Network Router Failure Prediction

---

Overview

This project is designed to estimate the failure probability of network routers using operational variables. By applying machine learning techniques, the system predicts potential failures, enabling proactive maintenance and reducing downtime in network infrastructures. The solution leverages supervised learning models, feature engineering, and calibration techniques to deliver reliable, actionable insights for network reliability teams.

The repository provides a comprehensive pipeline, from data preprocessing to model evaluation, including visualizations, performance metrics, and threshold optimization for decision-making.

---

Project Structure

network-router-failure-prediction/
├── data.py                  # Data loading, preprocessing, and feature engineering
├── main.py                  # Main pipeline: model training, evaluation, and output generation
├── outputs/                 # Generated visualizations, metrics, and analysis results
│   ├── 01_distributions_professional.png
│   ├── 02_correlation_matrix_professional.png
│   ├── 03_class_imbalance_professional.png
│   ├── 04_feature_target_relationships_professional.png
│   ├── model_comparison.csv
│   ├── calibration_curve_threshold.png
│   ├── gain_lift_charts.png
│   ├── cost_vs_threshold.png
│   ├── threshold_analysis_summary.csv
│   ├── calibration_random_forest
│   ├── calibration_xgboost.png
│   ├── prob_dist_xgboost.png
│   └── prob_dist_random_forest.png
│
├── data/                    # Processed datasets and feature metadata
│   ├── X_train_scaled.npy
│   ├── X_test_scaled.npy
│   ├── y_train.npy
│   ├── y_test.npy
│   └── feature_names.csv
│
├── models/                  # Serialized models and scalers
│   ├── scaler.pkl
│   ├── logistic_regression_best.pkl
│   ├── random_forest_best.pkl
│   ├── xgboost_best.pkl
│   ├── random_forest_sigmoid.pkl
│   ├── random_forest_isotonic.pkl
│   ├── xgboost_sigmoid.pkl
│   └── xgboost_isotonic.pkl
│
├── requirements.txt
└── datos_routers/

---

Objectives

• Predictive Modeling: Develop high-accuracy models to estimate router failure probability using operational variables.
• Feature Analysis: Identify key indicators of failure through exploratory data analysis and correlation studies.
• Model Calibration: Improve probability reliability using Platt Scaling and Isotonic Regression.
• Threshold Optimization: Balance false positives and negatives to minimize operational costs.
• Actionable Insights: Provide visualizations and metrics to support data-driven maintenance strategies.

---

Methodology

1. Data Preprocessing

• Feature Scaling: Standardization of operational variables using StandardScaler.
• Train-Test Split: Stratified splitting to maintain class distribution.
• Feature Selection: Analysis of feature importance and redundancy.

2. Exploratory Data Analysis (EDA)

• Distributions: Visualization of feature and target distributions.
• Correlation Matrix: Identification of multicollinearity and feature relationships.
• Class Imbalance: Assessment of failure/non-failure ratio and mitigation strategies.
• Feature-Target Relationships: Univariate analysis of operational variables vs. failure probability.

3. Model Training

• Logistic Regression: Baseline linear model for binary classification.
• Random Forest: Ensemble method for non-linear relationships and feature importance.
• XGBoost: Gradient-boosted trees for high performance and robustness.

4. Model Calibration

• Platt Scaling: Logistic regression-based calibration for sigmoid transformation.
• Isotonic Regression: Non-parametric calibration for monotonic probability adjustment.

5. Evaluation

• Metrics: Accuracy, Precision, Recall, F1-Score, ROC-AUC, and PR-AUC.
• Calibration Curves: Comparison of predicted vs. actual probabilities.
• Cost Analysis: Evaluation of misclassification costs at various thresholds.

6. Threshold Optimization

• Gain/Lift Charts: Assessment of model effectiveness in identifying high-risk routers.
• Cost vs. Threshold: Optimization of decision thresholds to minimize operational costs.

---

Models

Model	Description
Logistic Regression	Linear model for baseline performance.
Random Forest	Ensemble of decision trees for non-linear patterns.
XGBoost	Gradient-boosted trees for high accuracy and speed.
Random Forest (Calibrated)	RF with Platt Scaling or Isotonic Regression for reliable probabilities.
XGBoost (Calibrated)	XGBoost with Platt Scaling or Isotonic Regression for reliable probabilities.

---

Key Outputs

• Visualizations:
  • Feature distributions, correlation matrices, class imbalance, and feature-target relationships.
  • Calibration curves (before/after calibration) for all models.
  • Probability distributions of predicted failure probabilities.
• Metrics:
  • Model comparison CSV with performance metrics.
  • Threshold analysis summary for cost optimization.
• Charts:
  • Gain/lift charts for model effectiveness.
  • Cost vs. threshold analysis for decision-making.

---

How to Use

1. Clone the Repository:

   git clone https://github.com/your-username/network-router-failure-prediction.git
   cd network-router-failure-prediction

2. Install Dependencies:

   pip install -r requirements.txt

3. Prepare Data:

   • Place raw operational data in the datos_routers/ directory.
   • Ensure the data includes operational variables and failure labels.

4. Run the Pipeline:

   python main.py

   • This script executes:
     • Data loading and preprocessing (data.py).
     • Model training, calibration, and evaluation.
     • Generation of all outputs in the outputs/ directory.

5. Review Results:

   • Check the outputs/ directory for:
     • Visualizations (PNG files).
     • Performance metrics (CSV files).
     • Calibration and threshold analysis.

---

Dependencies

• Python 3.8+
• Key libraries:
  • pandas, numpy
  • scikit-learn, xgboost
  • matplotlib, seaborn
  • joblib, pickle

---

Author

This project is developed and maintained as part of an initiative to enhance network reliability through predictive analytics and machine learning.