Graph Neural Networks (GNN)

GNNs extend deep learning to graph-structured data (networks, molecules, knowledge graphs). Achieve 85-95% accuracy on node classification, link prediction, and graph classification - enabling social network analysis, drug discovery, and recommendation systems.

Why GNNs?

• Relational Data: Many problems involve entities and relationships
• Irregular Structure: Graphs don't fit into grids (unlike images)
• Message Passing: Learn from neighbors' features
• Better than Features: 10-30% improvement over hand-crafted features

Key Architectures

1. Graph Convolutional Networks (GCN)

• Convolutional layers for graphs
• Aggregate neighbor features
• Foundation for many GNN variants
• Good for node classification

2. GraphSAGE

• Sampling-based aggregation (scalable)
• Inductive learning (generalize to new nodes)
• Used in large-scale applications (Pinterest, Uber)

3. Graph Attention Networks (GAT)

• Attention mechanism for neighbor aggregation
• Learn importance of different neighbors
• State-of-the-art performance

4. Message Passing Neural Networks (MPNN)

• General framework for GNNs
• Nodes exchange messages with neighbors
• Flexible, many variants

Applications

Social Network Analysis

• Node Classification: Predict user attributes, communities
• Link Prediction: Friend recommendations (85-92% accuracy)
• Influence Propagation: Viral marketing, information diffusion
• Anomaly Detection: Fake accounts, bot detection

Drug Discovery

• Molecules as graphs (atoms = nodes, bonds = edges)
• Property prediction (toxicity, solubility)
• Drug-target interaction prediction
• 10-100x faster than traditional methods
• 90%+ accuracy on molecular property prediction

Knowledge Graphs

• Entity and relation embedding
• Link prediction (missing facts)
• Question answering over knowledge graphs
• Example: Google Knowledge Graph

Recommendation Systems

• User-item graph for recommendations
• Better than collaborative filtering (10-20% lift)
• Capture complex relationships
• Used by Pinterest, Alibaba, Netflix

Traffic & Logistics

• Road networks for traffic prediction
• Supply chain optimization
• Route planning

Tasks

Node Classification

• Predict label of each node
• Semi-supervised learning
• Example: Classify papers in citation network

Link Prediction

• Predict missing or future links
• Friend recommendations, knowledge completion
• 85-95% accuracy on many benchmarks

Graph Classification

• Classify entire graphs
• Example: Predict if molecule is toxic
• Pooling layers to aggregate graph-level features

Implementation

Libraries

• PyTorch Geometric: Most popular, 100+ GNN models
• DGL (Deep Graph Library): Scalable, TensorFlow/PyTorch
• Spektral: Keras-based GNN library
• NetworkX: Graph data structures

Process

1. Data: Construct graph (nodes, edges, features)
2. Model: Choose GNN architecture (GCN, GAT, GraphSAGE)
3. Train: Supervised/semi-supervised learning
4. Evaluate: Accuracy, F1, AUC

Best Practices

• Node Features: Rich features improve performance
• Graph Sampling: For large graphs (millions of nodes)
• Oversmoothing: Too many layers can hurt (2-3 layers often best)
• Inductive vs Transductive: New nodes? Use inductive (GraphSAGE)

Case Study: Social Network Friend Recommendations

• Scale: 10M users, 500M connections
• Model: GraphSAGE for link prediction
• Results:
  • Recommendation accuracy: 88% (vs 72% collaborative filtering)
  • Click-through rate: +42%
  • Engagement: +35%
  • Training time: 4 hours (vs 2 days traditional methods)

Pricing

• Small Graph (<1M nodes): ₹12-25L
• Large Graph (1-100M nodes): ₹35-70L
• Enterprise (100M+ nodes): ₹80L-2Cr