PubMed Graph Attention Network with Explainable AI

A comprehensive Graph Neural Network implementation with state-of-the-art explainability features for scientific literature classification

Project Information

Course Details

Course: Explainable AI
Instructor: Dr. Stefan Heindorf
Institution: Paderborn University
Semester: Summer 2025

Team Members

Mohammadparsa Rostamzadehkhameneh - Matriculation Number: 4038848
Alireza Rahnama - Matriculation Number: 4082518

Project Overview

This project implements a Graph Attention Network (GAT) for node classification on the PubMed citation network, focusing on diabetes-related scientific literature classification. The implementation features a comprehensive Explainable AI (XAI) framework that provides deep insights into model decision-making through attention pattern analysis, feature importance visualization, and multi-perspective explanations.

Classification Task

Dataset: PubMed Citation Network (Planetoid)
Task: Node classification for scientific papers
Classes:
- Diabetes Mellitus (Class 0)
- Experimental Diabetes (Class 1)
- Type 1 Diabetes (Class 2)
Graph Structure: 19,717 nodes, 108,365 edges, 500 TF-IDF features

Hardware Dependency Notice

Note: The results presented in this project were obtained on a Linux-based OS without GPU acceleration. If you run this code on a GPU or different hardware configuration, you may get different results.

Reported Results:

Original Model: ~72% accuracy
Optimized Model: ~75% accuracy
System: Linux CPU-based training

Important:

Results may vary significantly when running on GPU or different hardware setups.

Key Features

Advanced GAT Architecture

Multi-head attention mechanism with configurable attention heads
Sophisticated regularization: Dropout, Batch Normalization, Weight Decay
Stable training: Gradient clipping and early stopping mechanisms
Adaptive optimization: Learning rate scheduling with ReduceLROnPlateau

Comprehensive Explainability Framework

Attention Pattern Analysis: Visualize attention flow between papers
Feature Importance Analysis: Gradient-based feature attribution methods
Multi-head Attention Visualization: Compare specialization across attention heads
Class-specific Analysis: Understand feature differences between diabetes types
Network-level Insights: Subgraph attention pattern exploration

Advanced Visualization Suite

Interactive Network Graphs: NetworkX-based attention flow visualization
Multi-perspective Analysis: 6 different visualization types per analysis
Statistical Visualizations: Distribution analysis and comparative plots
Heatmap Representations: Attention strength matrices across nodes

Intelligent Feature Optimization

Dataset-wide Analysis: Importance evaluation across representative samples
Automatic Feature Selection: Cumulative importance-based filtering
Performance Optimization: Maintaining accuracy with reduced dimensionality
Comparative Framework: Before/after performance evaluation

Installation & Requirements

Setup (All Platforms)

# 1. Clone repository
git clone [your-repository-url]
cd pubmed-gat-explainable-ai

# 2. Create virtual environment
python -m venv venv

# 3. Activate environment
# Linux/Mac: source venv/bin/activate
# Windows: venv\Scripts\activate

# 4. Install dependencies
pip install -r requirements.txt
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv -f https://data.pyg.org/whl/torch-2.7.0+cpu.html

Run the Project

# Execute the complete pipeline
python pubmed_GNN.py

Usage Instructions

Quick Start (Complete Pipeline)

# Run the entire pipeline with one command
python pubmed_GNN.py

This executes the full pipeline:

Data preprocessing and validation
GAT model training with optimization
Comprehensive explainability analysis
Feature importance evaluation
Dataset optimization and retraining
Performance comparison and insights

Step-by-Step Execution

Step 1: Data Preprocessing

from pubmed_GNN import preprocess_pubmed_dataset

# Preprocess the PubMed dataset
processed_data, save_path = preprocess_pubmed_dataset('processed_data.pt')
print(f"Processed data saved to: {save_path}")

Step 2: Model Training

from pubmed_GNN import run_PubMed_Gat

# Train the GAT model
model_results = run_PubMed_Gat('./data/processed_data/processed_data.pt')
print(f"Model trained with test accuracy: {model_results['final_test_accuracy']:.4f}")

Step 3: Explainability Analysis

from pubmed_GNN import explain_gat_attention, explain_gat_features

# Attention pattern analysis
attention_results = explain_gat_attention(model_results, node_id=None)

# Feature importance analysis  
feature_results = explain_gat_features(model_results, top_k=20)

Step 4: Advanced Analysis

from pubmed_GNN import compare_class_features, analyze_dataset_for_optimization

# Class-specific feature comparison
class_results = compare_class_features(model_results, samples_per_class=5)

# Dataset-wide optimization analysis
optimization_analysis = analyze_dataset_for_optimization(model_results, num_samples=100)

Results & Key Insights

Model Performance

Metric	Original GAT	Optimized GAT	Improvement
Test Accuracy	72.1%	75.8%	+3.7%
Validation Accuracy	75.0%	78.8%	+3.8%
Feature Count	500	218	56.4% reduction
Model Parameters	48,777	21,705	55.5% reduction
Train-Val Gap	0.250	0.212	Better generalization
Importance Retained	100%	80%	Minimal loss
Training Epochs	34 (early stop)	87 (early stop)	More stable convergence

Explainability Insights

Attention Pattern Discoveries

Head Specialization: Different attention heads focus on distinct types of paper relationships
Citation Influence: Strong attention weights on highly-cited diabetes research papers
Class-specific Patterns: Each diabetes type shows unique attention distribution patterns
Local vs Global: Balance between local neighborhood and global graph structure utilization

Feature Importance Findings

Sparse Feature Landscape: Only 22 out of 500 features (4.4%) showed significant importance (>0.01)
No High-Impact Features: Zero features exceeded high importance threshold (>0.05)
Top Feature Importance: Maximum importance score of 0.0269 indicates highly distributed information
Noise Identification: 282 features removed (56.4% reduction) with performance improvement
Feature Distribution: Importance evenly spread across remaining features, suggesting ensemble-like behavior

Optimization Impact

Dimensionality Reduction: 500 → 180 features (56% reduction)
Maintained Performance: 80% importance retention with improved accuracy
Computational Efficiency: Significant reduction in training time and memory usage
Generalization: Improved model robustness through noise reduction

Technical Architecture

GAT Model Architecture

Input Layer: [500 TF-IDF features]
↓ [Input Dropout: 0.25]
Multi-Head GAT Layer 1: [32 × 3 heads = 96 dimensions]
↓ [Batch Normalization]
↓ [ReLU Activation]
↓ [Dropout: 0.5]
Single-Head GAT Layer 2: [3 output dimensions]
↓ [Softmax for classification]
Output: [3 diabetes classes]
Parameters: 48,777 (Original) → 21,705 (Optimized)

Explainability Pipeline

Model Predictions
    ↓
Attention Weight Extraction → Attention Pattern Analysis
    ↓                              ↓
Gradient Computation → Feature Importance → Visualization Suite
    ↓                              ↓
Subgraph Analysis → Network Visualization → Insights Generation

Data Flow Architecture

Raw PubMed Data → Preprocessing → Feature Scaling → GAT Training
                     ↓               ↓              ↓
              Validation Split → Edge Processing → Model Evaluation
                     ↓               ↓              ↓
              XAI Analysis ← Feature Selection ← Performance Analysis

Resources & References

Core Papers & Methods

Graph Neural Networks

Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y. (2018). "Graph Attention Networks." International Conference on Learning Representations (ICLR).
- Implementation: Core GAT architecture, multi-head attention mechanism

Dataset & Graph Construction

Sen, P., Namata, G., Bilgic, M., Getoor, L., Gallagher, B., Eliassi-Rad, T. (2008). "Collective Classification in Network Data." AI Magazine, 29(3), 93-106.
- Implementation: PubMed dataset structure, citation network construction

Explainable AI & Feature Attribution

Adadi, A., Berrada, M. (2018). "Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)." IEEE Access, 6, 52138-52160.
- Reference: XAI theoretical foundations and survey of explainability methods
Sundararajan, M., Taly, A., Yan, Q. (2017). "Axiomatic Attribution for Deep Networks." International Conference on Machine Learning (ICML), pp. 3319-3328.
- Implementation: Gradient-based feature importance calculation (Gradient×Input method)
Yuan, H., Yu, H., Gui, S., Ji, S. (2023). "Explainability in Graph Neural Networks: A Taxonomic Survey." IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5), 5782-5799.
- Reference: Comprehensive survey of GNN explainability methods

Graph Neural Network Explainability

Ying, R., Bourgeois, D., You, J., Zitnik, M., Leskovec, J. (2019). "GNNExplainer: Generating Explanations for Graph Neural Networks." Advances in Neural Information Processing Systems (NeurIPS), pp. 9240-9251.
- Inspiration: Explainability framework design, attention-based explanations
Vig, J., Belinkov, Y. (2019). "Analyzing the Structure of Attention in a Transformer Language Model." Proceedings of the 2019 ACL Workshop BlackboxNLP, pp. 63-76.
- Reference: Attention analysis and visualization methodologies

Feature Selection & Optimization

Guyon, I., Elisseeff, A. (2003). "An Introduction to Variable and Feature Selection." Journal of Machine Learning Research, 3, 1157-1182.
- Implementation: Feature selection methodology and evaluation metrics

Technical Libraries & Frameworks

Fey, M., Lenssen, J.E. (2019). "Fast Graph Representation Learning with PyTorch Geometric." ICLR Workshop on Representation Learning on Graphs and Manifolds.
- Implementation: Graph data structures, GAT layers, dataset loading

Medical Data & Applications

Johnson, A.E., Pollard, T.J., Shen, L., et al. (2016). "MIMIC-III, a freely accessible critical care database." Scientific Data, 3, 160035.
- Reference: Medical data processing and healthcare AI applications

AI Usage Acknowledgment

In accordance with academic transparency requirements, the following sections utilized AI assistance during development:

Code Implementation

Data Preprocessing Pipeline (AI-assisted): Complex data validation, feature scaling methodology, and error handling patterns
Visualization Functions (AI-assisted): Matplotlib configuration, NetworkX graph layouts, color schemes, and multi-subplot arrangements
Statistical Analysis Methods (AI-assisted): Feature importance aggregation, distribution analysis, and comparative statistics

GAT Model Optimization

Hyperparameter Tuning (AI-assisted): Suggestions for optimal learning rates, dropout rates, hidden dimensions, and number of attention heads
Advanced Training Techniques (AI-assisted): Implementation of gradient clipping, learning rate scheduling, early stopping, and batch normalization integration
Performance Enhancement Methods (AI-assisted): Suggestions for attention dropout, weight decay optimization, and multi-layer architecture design

Visualization Design

Attention Network Plots (AI-assisted): NetworkX layout algorithms, edge styling, node positioning, and interactive elements
Statistical Visualizations (AI-assisted): Histogram configurations, heatmap color maps, and subplot arrangements
Comparative Analysis Charts (AI-assisted): Bar chart designs, legend configurations, and annotation placement

Analysis Framework

Feature Selection Logic (AI-assisted): Cumulative importance calculation and threshold determination
Explainability Framework Design (AI-assisted): Multi-perspective analysis approach and attention pattern interpretation methods.

Project Structure

pubmed-gat-explainable-ai/
├── 📄 pubmed_GNN.py                   # Main implementation file
├── 📄 README.md                       # This documentation
├── 📁 data/
│   ├── 📁 PubMed_data/                # Raw dataset directory (auto-created)
│   └── 📁 processed_data/             # Processed datasets
│       ├── processed_data.pt          # Initial processed data
│       └── feature_selected_data.pt   # Feature-optimized dataset
└── 📁 models/
    └── GAT_Model.pt                   # Trained model checkpoint

⬆ Back to Top

Name		Name	Last commit message	Last commit date
Latest commit History 10 Commits
data/processed_data		data/processed_data
models		models
README.md		README.md
Report.pdf		Report.pdf
pubmed_GNN .py		pubmed_GNN .py
requirements.txt		requirements.txt

Alirezarahnamaa/PubMed_GNN_Project

Folders and files

Latest commit

History

Repository files navigation