Development and Evaluation of a Novel Framework Integrating LLMs, GNNs, and RL for Network Traffic Prediction and Optimization

Overview

This project provides an end-to-end research framework integrating:

1. LLM-based semantic parsing of network traffic logs.
2. Graph Neural Network (PyTorch Geometric) for traffic load prediction.
3. Reinforcement Learning (Stable-Baselines3) for routing / resource optimization.
4. Integrated loop for adaptive decision making.

Key Features

• Modular architecture (src/llm, src/gnn, src/rl, src/pipeline).
• Synthetic data generation and hooks for real datasets (CAIDA, NSL-KDD).
• Config-driven experimentation via configs/config.yaml.
• Reproducible training scripts.
• Evaluation metrics and visualization utilities.

Installation

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Synthetic Data Generation

python data/synthetic/generate_synthetic.py --num-nodes 12 --timesteps 500

Outputs:

• data/processed/traffic_timeseries.parquet
• data/processed/topology_edges.csv
• data/processed/logs.jsonl

Train GNN

python src/gnn/train_gnn.py --config configs/config.yaml

Train RL Agent

python src/rl/train_rl.py --config configs/config.yaml

Full Pipeline Loop

python src/pipeline/integrated_loop.py --config configs/config.yaml

Evaluation

Artifacts:

• Metrics JSON in outputs/reports/
• Plots in outputs/figures/

Real Datasets

You may integrate:

• CAIDA Anonymized Internet Traces (requires agreement)
• NSL-KDD (intrusion detection semantics)
• MAWI traffic archives
• RIPE Atlas ping/latency for augmentation

Convert raw data into:

• traffic_timeseries.parquet (columns: time, src, dst, bytes, packets, protocol, ...)
• topology_edges.csv (src, dst, capacity)
• logs.jsonl (raw textual lines for LLM parsing)

Research Extensions

• Replace static GNN with temporal architectures (TGN, DCRNN).
• Multi-objective RL (Pareto frontier of latency vs energy).
• Graph-level policy networks with graph attention policy encoding.
• Continual learning for concept drift.

Reproducibility

• Seed control in config.
• Deterministic PyTorch flags (note: full determinism may degrade GPU performance).
• Logged environment + model hyperparameters.

Citation (Template)

Odoh, J. (2025). Development and evaluation of a novel framework integrating LLMs, GNNs, and RL for network traffic prediction and optimization. Retrieved October 18, 2025

License

Apache 2.0 (adjust as needed).