Federated Learning for Intelligent Transportation Systems

Overview

This project presents a privacy-preserving traffic forecasting framework using Federated Learning (FL) and Spatio-Temporal Graph Neural Networks (STGNNs). The objective is to accurately predict future traffic flow while preserving data privacy by ensuring that raw traffic sensor data remains localized on client devices and is never shared with a central server.

Two federated deep learning architectures are implemented and compared:

1. GNN-GRU-KAN

• Graph Attention Network (GAT)
• Graph Convolutional Network (GCN)
• Gated Recurrent Unit (GRU)
• Kolmogorov-Arnold Network (KAN) Prediction Head

2. STGAT+GCN

• Graph Attention Network (GAT)
• Graph Convolutional Network (GCN)
• Gated Recurrent Unit (GRU)
• Fully Connected Prediction Head

Experiments were conducted using the California PeMS Traffic Dataset under a federated learning environment utilizing the FedAvg aggregation algorithm.

---

Key Features

• Federated Learning using FedAvg
• Privacy-preserving traffic forecasting
• Graph-based spatial feature extraction
• GRU-based temporal sequence modeling
• KAN and Fully Connected prediction heads
• Multi-step traffic flow forecasting
• Comparative evaluation of federated architectures
• Comprehensive performance analysis using multiple forecasting metrics

---

Dataset

California PeMS Traffic Dataset

The experiments utilize the California Performance Measurement System (PeMS) dataset:

• Real-world traffic sensor data
• 15-minute sampling interval
• 435 traffic sensors
• Multi-city California traffic network
• Large-scale spatio-temporal traffic observations

Dataset Access

Due to GitHub file size limitations, the complete dataset and project resources are hosted on Google Drive.

Download Link

🔗 Project Data Archive

https://drive.google.com/file/d/1iTa07LPK1uC2051IQJ6UFwdVm_B4M9jy/view?usp=drive_link

Please download and extract the archive before running the notebooks or training scripts.

Data Split

Dataset Split	Percentage
Training	70%
Validation	15%
Testing	15%

Forecasting Configuration

Parameter	Value
Lookback Window	12 Timesteps
Forecast Horizon	5 Timesteps
Sampling Interval	15 Minutes

---

Model Architecture

Shared Spatio-Temporal Encoder

Traffic Data
     │
     ▼
Graph Attention Network (GAT)
     │
     ▼
Graph Convolutional Network (GCN)
     │
     ▼
GRU Encoder
     │
     ▼
Layer Normalization
     │
     ▼
Prediction Head

GNN-GRU-KAN

GAT → GCN → GRU → LayerNorm → KAN

KAN Configuration

• Hidden Layers: [128, 128, 64]
• Grid Size: 8
• Spline Order: 3

STGAT+GCN

GAT → GCN → GRU → LayerNorm
                    │
                    ▼
      Linear → SiLU → Dropout → Linear

---

Federated Learning Configuration

Parameter	Value
Clients	4
Aggregation Algorithm	FedAvg
Communication Rounds	100
Optimizer	AdamW
Loss Function	Huber Loss
Batch Size	64
Gradient Clipping	1.0
Parameter Clamping	[-2, 2]

---

Training Hyperparameters

Hyperparameter	Value
Learning Rate	1e-3
Weight Decay	1e-4
GRU Hidden Size	128
GRU Layers	2
GRU Dropout	0.2
FC Dropout	0.1
KAN Grid Size	8
KAN Spline Order	3

---

Experimental Results

Final Test Performance

Metric	GNN-GRU-KAN	STGAT+GCN
MAE	5.3474	3.7010
MSE	44.3993	21.9106
RMSE	6.6633	4.6809
MAPE (%)	12.4639	7.9356
SMAPE (%)	10.9100	7.5799
R² Score	0.9588	0.9797

Key Findings

The STGAT+GCN model consistently outperformed the GNN-GRU-KAN architecture across all evaluation metrics.

Possible reasons include:

• Greater stability during FedAvg aggregation
• Lower parameter complexity
• Better compatibility with federated averaging
• Reduced sensitivity to client-side model divergence
• Improved convergence behavior across communication rounds

---

Project Structure

.
├── notebooks/
│   └── Finalized_Model_GRU.ipynb
│
├── results/
│   ├── metrics/
│   └──plots/
│
├── requirements.txt
└── README.md

---

Installation

Clone the repository:

git clone https://github.com/SabudhFoundation/federated-traffic-forecasting.git
cd federated-traffic-forecasting

Download the dataset from: https://drive.google.com/file/d/1iTa07LPK1uC2051IQJ6UFwdVm_B4M9jy/view?usp=drive_link

Install dependencies:

pip install -r requirements.txt

---

Dependencies

• Python 3.12.x
• PyTorch
• PyTorch Geometric
• NumPy
• Pandas
• scipy
• networkx
• Scikit-learn
• Matplotlib
• efficient-kan

---

Running the Project

Launch the notebook:

jupyter notebook Finalized_Model_GRU.ipynb

Update the Dataset path in notebook Finalized_Model_GRU " CSV_PATH = "Path on your system" "

---

Evaluation Metrics

The following metrics are used to evaluate forecasting performance:

• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• Mean Absolute Percentage Error (MAPE)
• Symmetric Mean Absolute Percentage Error (SMAPE)
• Coefficient of Determination (R²)

---

Future Work

• Traffic-aware federated aggregation strategies
• FedProx implementation
• Asynchronous Federated Learning
• Non-IID client partitioning using clustering
• Experiments on METR-LA and additional traffic datasets
• Enhanced aggregation methods for KAN layers
• Real-world deployment on edge and IoT devices

---

Authors

• Manav Sharma
• Justin Johnson
• Aryan Raj
• Revanth Badithabonu
• Gajavada Sanjeevkumar

Mentors

• Dr. Sukhjit Singh Sehra
• Mr. Sanchit Umate

---

License

This project was developed as part of the Sabudh Foundation Data Science Internship Program and is intended for academic and research purposes.