Yelp Restaurant Recommendations with GNNs

Project Overview

This project implements a recommendation system for restaurants using the Yelp dataset and Graph Neural Networks (GNNs). The system leverages heterogeneous graph neural networks to model user-restaurant interactions and make personalized restaurant recommendations.

Features Implemented

• Heterogeneous Graph Construction - Built from user-restaurant interactions and metadata
• Multiple GNN Models - Implemented baseline GraphSAGE, improved model with GAT and Transformer layers, and hard sampling model
• Comprehensive Evaluation - Measured performance using Recall@K, Precision@K, MAP, and AUC
• Comparative Analysis - Benchmarked different loss functions (BPR vs. BCE) and model architectures

Implementation Details

The recommendation system includes several key components:

1. Data Processing Pipeline - Processes raw Yelp JSON data into a heterogeneous graph
2. Model Architectures:
   • Baseline: GraphSAGE-based GNN with simple edge prediction
   • Improved: Enhanced GNN with GAT, Transformer layers, and skip connections
   • Hard Sampling: Improved model with hard negative sampling
3. Training Framework - Supports different loss functions and evaluation metrics
4. Evaluation Suite - Comprehensive testing and performance analysis tools

Learning Outcomes

This project allowed me to:

• Gain expertise in implementing GNNs for recommendation systems
• Understand the challenges of processing large-scale heterogeneous graphs
• Develop skills in comparative model analysis and evaluation
• Implement advanced training techniques like hard negative sampling

GitHub Repository

The complete source code is available on my GitHub repository: github.com/apratim-mishra/gnn_yelp