This repository contains the works for ST5188 Statistical Research Project.
ABOUT:
In this project, we propose novel approaches to alleviate over-smoothing in LightGCN to enhance its performance in graph representation learning for recommender systems. By exploring solutions such as cluster-based sampling and dimensionality reduction, our aim is to enhance the performance of LightGCN in graph representation learning for recommender systems, while addressing the limitations and challenges posed by over-smoothing.
This project was completed in fulfillment of the capstone project module, ST5188 Statistical Research Project, as part of the NUS Statistics M.Sc. by Coursework programme.
CONTENTS:
- Cluster-LightGCN (Cluster-Based Sampling)
Self-contained code in Jupyter notebook format for running our lightweight implementation of LightGCN over the MovieLens100k dataset, and for testing the Cluster-based sampling approach.
- SimpleLGN (Collapsing User-Item Representation)
This is the modified version of the Original LightGCN source code, for testing simplegcn and mixgcf+simplegcn.
- Tuned-MixGCF (Improving Training)
This is the code for testing the tuned MixGCF approach.
- Experimental Files
This folder contains all intermediate experimental jupyter notebooks.
- Reports
This folder contains all intermediate and final reports, and final presentation materials prepared over the course of this project.