This repository contains the AIMS (Adaptive and Intelligent Management of Slicing) framework for classifying how network slicing policies impact Intelligent Transportation Systems (ITS) applications. The framework uses machine learning to analyze Quality of Service (QoS) metrics and predict impact levels, enabling dynamic resource allocation in vehicular networks.
- Multi-model approach: Implements three complementary classifiers with different strengths
- Random Forest: Tree-based ensemble with bagging for robust baseline performance
- TabNet: Deep learning with attention mechanisms optimized for tabular data
- CatBoost: Gradient boosting with native categorical feature support and advanced regularization
- Temporal-aware validation: Uses GroupKFold cross-validation (5 splits) to prevent data leakage in time-series data
- Feature engineering: 15+ engineered features from core network metrics
- Class-balanced training: Handles imbalanced impact level distribution across all models
- Automated hyperparameter optimization: Uses Optuna for efficient parameter tuning
- Comprehensive evaluation: Generates confusion matrices, feature importance, and comparative analysis
The framework processes vehicular network QoS data derived from the experimental setup described in T. do Vale Saraiva et al., "An Application-Driven Framework for Intelligent Transportation Systems Using 5G Network Slicing," IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 8, pp. 5247–5260, Aug. 2021 (based on saraivacode/framework_its_sdn)
The dataset includes:
- 158 vehicles traveling on urban roads with QoS data from up to 15 vehicles communicating simultaneously
- 450 seconds of network measurements
- 4 application classes: Safety (S), Efficiency (E), Entertainment (E2), Generic (G)
- Core metrics: RTT (latency), PDR (packet delivery ratio), Throughput and over 15 engineered features
# Python Python3.12.4 or higher
python --version
# Install required packages
pip install -r requirements.txtcatboost==1.2.8
joblib==1.5.1
matplotlib==3.10.3
numpy==2.3.2
optuna==4.4.0
pandas==2.3.1
pytorch_tabnet==4.1.0
scikit_learn==1.7.1
seaborn==0.13.2
tabulate==0.9.0
torch==2.6.0+cu124
code/
├── main.py # Main training pipeline
├── train_model_catboost.py # CatBoost training script
├── train_model_rf.py # Random Forest training script
├── train_model_tabnet.py # TabNet training script
├── preprocess_dataset.py # Data preprocessing utilities
├── impact_labeling.py # Impact level assignment logic
├── save_utils.py # Artifact saving utilities
├── compare_results.py # Compare output results script
data/ # Dataset directory
└── aims_dataset.csv
other/
├── requirements.txt # Python dependencies
├── README.md # This file
Train all three models with 15 trials:
python main.py --compare --csv ../data/aims_dataset.csv --n-trials 15 --n-trials-tabnet 15Train specific models with custom parameters:
# Custom dataset with increased optimization trials
python main.py --csv ./data/custom_dataset.csv --n-trials 100
# Skip computationally expensive model and generate comparison
python main.py --skip-tabnet --compare
# Development mode - test comparison logic only
python main.py --test-comparison-only
# Targeted training with custom output directory
python main.py --skip-rf --results-dir ./experiments/run_001
# Random Forest
python main.py --compare --csv ../data/aims_dataset.csv --n-trials 15 --skip-catboost --skip-tabnet
# CatBoost
python main.py --compare --csv ../data/aims_dataset.csv --n-trials 15 --skip-rf --skip-tabnet
# TabNet
python main.py --compare --csv ../data/aims_dataset.csv --n-trials-tabnet 15 --skip-catboost --skip-rfDataset & Core Configuration:
--csv: Path to the dataset CSV file (default: ../data/aims_dataset.csv)--random-state: Random seed for reproducibility (default: 42)--results-dir: Base directory for storing model results (default: ../results)
Cross-Validation & Optimization:
--n-splits: Number of GroupKFold cross-validation splits (default: 5)--n-trials: Number of Optuna optimization trials for RandomForest and CatBoost (default: 40)--n-trials-tabnet: Number of Optuna trials for TabNet (default: 40)
Model Selection:
--skip-rf: Skip Random Forest training--skip-tabnet: Skip TabNet training--skip-catboost: Skip CatBoost training
Analysis & Testing:
--compare: Generate comparison report after training completion--test-comparison-only: Skip training and run only comparison logic
The framework uses a weighted-average approach to assign impact levels based on application-specific QoS thresholds. Each metric (RTT, PDR, throughput) is scored 0-3, then combined using application-specific weights.
Impact Levels:
- 0 (Adequate): Adequate performance, all QoS requirements met
- 1 (Warning): Slight degradation, non-critical applications affected
- 2 (Severe): Significant degradation, multiple applications impacted
- 3 (Critical): Severe degradation, safety applications at risk
Application Weights:
| Application | Latency | Loss | Throughput | Priority |
|---|---|---|---|---|
| Safety (S) | 0.5 | 0.3 | 0.2 | Critical |
| Efficiency (E) | 0.3 | 0.4 | 0.3 | High |
| Entertainment (E2) | 0.2 | 0.3 | 0.5 | Medium |
| Generic (G) | 0.3 | 0.3 | 0.4 | Low |
The framework generates 15+ features from core metrics including temporal features (rolling mean/std with 3-sample window, rate of change), derived metrics (loss ratio, throughput utilization), and categorical data (one-hot encoded application categories).
The framework generates comprehensive outputs for analysis and deployment:
Model Files: Trained models saved in pickle format for deployment and inference
Performance Analysis: Training results with metrics, confusion matrices (both raw and normalized), and feature importance visualizations
Optimization Reports: Hyperparameter optimization history and training curves for model tuning analysis
Data Exports: Processed datasets, group classifications, and class weights for reproducibility
Comparative Analysis: Model comparison results and consolidated performance metrics across all approaches
If you use this code in your research, please cite:
@article{saraiva2025aims,
title={AIMS: Adaptive and Intelligent Management of Slicing for Next-Generation ITS Networks},
author={Saraiva, Tiago do Vale},
journal={},
year={2025},
publisher={}
}This project is licensed under the MIT License.
- Federal University of State of Rio de Janeiro (UNIRIO)
- Dataset based on saraivacode/framework_its_sdn, which uses:
- Tiago do Vale Saraiva - [email protected]