Recommender system using GAT and GraphSAGE

In my Spring 2023 semester at Georgia Tech, I took the CSE 6240 Web Search and Text Mining class. For the final project, I worked with a team of 3 other students to build a recommender system using GraphSAGE and Graph Attention Network.

Product recommendation system is one of the most popular applications of link prediction. This project reviews various approaches to the link prediction task, which can be broadly described in two steps: creating embeddings/feature vectors to represent each node and using the node embedding to predict the probability of a link between two nodes. To best apply link prediction tasks to real-world data, we use the Instacart Market Basket Analysis dataset from Kaggle. In this paper, we implement four established link prediction methods to recommend products to users in the Instacart dataset: two basline approaches (Matrix Factorization and Node2vec), as well as two more novel approaches (GraphSAGE and Graph Attention Network). We find that GraphSAGE performs the best of the four approaches, achieving an Area-Under-the-Curve score of 0.952.

Project report

Project presentation