This is the official repository of "EDGE-Rec: Efficient and Data-Guided Edge Diffusion for Recommender System Graphs", submitted as final project for the CMU course 10-708: Probabilistic Graphical Models.
We train and evaluate the model on the ML-100k dataset. We construct a custom 90-10 train-test split of the edges by adopting a stratified sampling approach to ensure that each user is represented in both the training and validation split.
We train on 1000 diffusion steps for 10000 iterations on a single A100 GPU in the Google Colab environment with batch size 16, patch size 50.
Results can be replicated in a step-by-step fashion by running the execute.ipynb notebook.
The denoising diffusion model borrows from denoising-diffusion-pytorch.