Diffusion Recommender Model - Reproduction and Extension

This is a reproduction and extension of the paper at SIGIR 2023:

Diffusion Recommender Model

Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, Tat-Seng Chua

Our reproduction is based on the original code. We extend the original code to include the following features:

• WandB logging
• learnable temporal weighting feature and visualization
• more suitable early stopping with increased patience
• additional clustering algorithms for L-DiffRec and LT-DiffRec

More details can be found in the report, for a quick overview see the poster.

Getting Started

We provide our working conda environment in the environment.yml file and add instructions to obtain the trained models.

1. Clone this repo if not already done so.

2. set up environment using conda and the provided environment.yml file:

   conda create --name rs --file environment.yml

3. activate the environment:

   source activate rs

4. Download the checkpoints released on Google drive.

   gdown --id 1bPnjO-EzIygjuvloLCGpqfBVnkrdG4IH
   unzip checkpoints.zip -d checkpoints
   rm checkpoints.zip

5. Choose whether to enable wandb logging:

   If you do not want to use WandB logging, set the following environment variable:

   export WANDB_MODE=disabled

   To use WandB logging, set your API key:

   We use to log the training process. To use it, you need to create a new project in WandB and get your API key. Then run the following command, and provide your API key:

   python3 -m wandb login

Addditionally, we provide a create_env.job file to create the conda environment and install all other requirements on a cluster using slurm.

Usage

WandB Logging

• project name (entity) needs to be adapted in main.py for each model in the wandb.init() function.
• --run_name: set flag in order to add a description to the run name, e.g. --run_name=TEST

Disabling WandB Logging

If you do not wish to use WandB logging, set the following environment variable:

export WANDB_MODE=disabled

Data

The experimental data are in the 'datasets' folder, and include Amazon-Book, Yelp and MovieLens-1M.

Checkpoints

The checkpoints released by the authors are in './checkpoints' folder. To download them follow instructions in the Getting Started section.

Run our Modifications

We use argparse flags to modify the hyperparameters of the models:

• modify early stopping patience: add flag --patience to set the number of epochs to wait before early stopping, default is 20. Motivation for this change is that the original value was not chosen optimal.

• use learnable temporal weighting feature: add flag --mean_type=x0_learnable to use learnable temporal weighting feature.

• to use the STAMP-like weighting in T-DiffRec: add the flag --attention_weighting

• visualize the weights using m-PHATE directly integrated into main.py: add flags --visualize_weights and --mean_type=x0_learnable. The generated graphs are saved on WandB. Alternatively, it is also possible to use the saved numpy array in the mPHATE folder to generate the graphs locally by passing only the flag --mean_type=x0_learnable and then use the visualize_weights.py script in a second step.

  --workers specifies the number of parallel processes to use for the mPHATE algorithm.

  python utils/visualize_weights.py --dataset=amazon-book_clean --run_name=learnable_weight_test --model_type=T-DiffRec --seed 1 --workers 10

For L-DiffRec and LT-DiffRec:

• modify cluserting algorithm: add flag --clustering_method=kmeans' to use default clustering method kmeans. Other implementations are hierarchical, gmm. -- modify the amount of clusters used: add flag --n_cate 2` to set the number of clusters, default is 2.

Run the Reproduction

The reproductions can be run using the provided .job files. Here we include example commands only. Default values are chosen to match the original paper's choices as far as it was possible to infer them.

DiffRec

python ./DiffRec/main.py --data_path ./datasets/ --dataset=yelp_clean --cuda --gpu=1 

For reproduction of the original paper, we used the commands in the run_DiffRec.job file.

L-DiffRec

python ./L-DiffRec/main.py --data_path ./datasets/ --dataset=yelp_clean --cuda --gpu=1 

For reproduction of the original paper, we used the commands in the run_L.job file.

T-DiffRec

python ./T-DiffRec/main.py --data_path ./datasets/ --dataset=yelp_clean --cuda --gpu=1 

For reproduction of the original paper, we used the commands in the run_T.job file.

LT-DiffRec

python ./LT-DiffRec/main.py --data_path ./datasets/ --dataset=yelp_clean --cuda --gpu=1 

For reproduction of the original paper, we used the commands in the run_LT.job file.

Run Inference Only

We assume checkpoints have been downloaded and are in the checkpoints folder. Otherwise follow instructions in the "Getting Started" section.

Example Usage for Inference

python ./DiffRec/inference.py --data_path ./datasets/ --dataset=yelp_clean --cuda --gpu=1 

More examples of inference commands can be found in the linked .job files too.