Graph-based-Recommendation-System

Structure of our repository

.
|-- README.md (here)
|-- log (saved tensorboard log file)
|-- ml-100k (raw dataset)
|-- weights (saved model weights)
|-- text (grid search text and ablation study results)
|-- data (processed data)
|-- Network Visualization.ipynb (Notebook for visualization of graph)
|-- scripts
  -- run.sh(shell scripts for running gridsearch experiments)
  -- dataset.py 
  -- train.py 
  -- model.py
  -- utils.py
  -- loss.py

Introduction of our project

In this project, we use the link prediction based on the bipartite graph that represents therelationship between the user and item. The message passing by graph convolution allows us to describe users using items’ information, and vice versa. The rating prediction is forced to fit the user graph and item graph, such that the normalizedLaplacian is used to compute the Dirichlet norms as the regularization term. The experiment has shown that our model can reach a lowest RMSE around 0.938. Even without any use of side features, the RMSE can still stay below 0.95.

Prerequiste of packages for our model:

python==3.7
pandas==0.24.2
matplotlib==2.2.2
argparse==1.1
tqdm==4.31.1
tensorboardX==1.7
numpy==1.17.3
torch==1.3.1
networkx==2.4
folium==0.10.1

Running the model

python3 scripts/train.py --rate_num 5 \
			--lr 0.01 \
			--weight_decay 0.00001 \
			--num_epochs 1000 \
			--hidden_dim 5 \
			--side_hidden_dim 5 \
			--out_dim 5 \
			--drop_out 0.0 \
			--split_ratio 0.8 \
			--save_steps 100 \
			--log_dir './log' \
			--saved_model_folder './weights' \
			--dataset_path './ml-100k' \
			--save_processed_data_path './data' \
			--use_side_feature 1 \
			--use_data_whitening 1 \
			--use_laplacian_loss 1 \
			--laplacian_loss_weight 0.05

You can observe the loss curve through the training by runing:

tensorboard --logdir=log/name_of_your_saved_file

Figure 1: Loss curve.

Gridsearch for best hyperparameters by running shell script:

The reult(RMSE) will be saved in a gridsearch.txt under the text folder.