Research project / arXiv preprint

Graph neural networks for product recommendation.

This project studies product recommendation on the Amazon co-purchase graph under link prediction settings, comparing LightGCN, GraphSAGE, GAT, and PinSAGE across model performance, training complexity, scalability, and generalization trade-offs.

Read Preprint View Results

Method

The study frames recommendation as a graph link prediction problem, using co-purchase edges to learn product-product relationships and compare both general-purpose and recommendation-specific GNN architectures.

LightGCN GraphSAGE GAT PinSAGE Link Prediction
01

Graph Setup

Products are represented as graph nodes, with co-purchase relationships used as recommendation signal.

02

Model Comparison

LightGCN, GraphSAGE, GAT, and PinSAGE are evaluated under a shared recommendation task.

03

Training Pipeline

Training scripts and notebooks support model-specific optimization, logging, and result analysis.

04

Evaluation

Results focus on performance, scalability, training complexity, and deployment-oriented trade-offs.

Results

The result plots summarize comparative performance and training behavior across the tested GNN models.

LightGCN training history curve

LightGCN Training

Training history for the recommendation-specific graph collaborative filtering model.

GAT training and validation curves

GAT Training

Training and validation behavior for the graph attention model.

GraphSAGE training curve

GraphSAGE Training

Training curve for the neighborhood aggregation baseline.

PinSAGE training curve

PinSAGE Training

Training dynamics for the recommendation-focused graph model.