* Graph Recommender System for Amazon Product Reviews
Summary
Traditional recommendation systems include ensemble methods, matrix factorization, and feature-based methods, to name a few. We wanted to explore a novel approach to recommendation systems by using graph neural networks. Which makes sense intuitively since products are related to each other through reviews; a natural representation of the data is a graph. We used the Amazon 2018 Product Review dataset, and conducted link regression to predict user reviews.
Brief details
- Developed a Graph Neural Network based recommendation system for Amazon data to better capture customer relationships
- Integrated word embeddings using the all-MiniLM-L6-v2 model from HuggingFace to generate more accurate node embeddings
- Compared models such as GraphSAGE, GAT, and GCN achieving an RMSE of 1.08, surpassing traditional KNN and SVD methods