Handling New Users and Items: A Comparative Study of Inductive Recommenders

Requirements

The following applications must be installed on your system.

python3.10
python3.10-venv
docker
docker-compose-plugin

We note this project requires more than 200GB of disk space for the Yelp KG construction.

FIRST GET APPROVAL FOR DATASET USAGE, SEE datasets/README.md

Instantiate the environment and preprocess datasets

If above requirements are met, you can run the following commands to install the environment and preprocess the datasets. Assuming you are in the root directory of the project you can run the docker_builder.sh script to build the docker image. It will build and run required files. It will prompt for download of the Yelp Dataset, such that it can be processed. After downloading and placing in the required folder, run the script again.

./docker_builder.sh

After running, a terminal will at /app in the docker container. You can run the following commands to preprocess the datasets.

cd datasets; ./dataset_preparations.sh

You may need to run chmod +x dataset_preparations.sh to make the script executable.

Training and evaluation

Run the following command to start the tuning of the methods. You can specify the GPU indices as --gpu 0 0 1 1 to use multiple GPUs or same GPU multiple times.

python3 py_run_parallelized.py

After tuning, run to fully train given the parameters above.

python3 py_run_multi.py

Rerun using --eval to evaluate the models.

python3 py_run_multi.py --eval

Computing metrics

To compute the metrics, run the following command. It will compute the metrics for all methods and datasets.

cd evaluate
./eval_script.sh

There are now multiple pickle files, one for each dataset, method, and experiment. Loading these files will give all user metrics with k's from 1 to 50. Perform whatever analysis you want on these files.

General information about the framework

Under configuration you will find all parameters for the models as well as dataset and experiment configuration. For datasets you can define a mapping of ratings, such as explicit to implicit and apply filters, such as minimal number of ratings using CountFilter:

ml_mr_1m = DatasetConfiguration('ml-mr-1m', lambda x: {x > 3: 1}.get(True, 0),
                                filters=[CountFilter(lambda c: c >= 5, sentiment=Sentiment.POSITIVE)],
                                max_users=12500)

All methods, experiments and datasets must be defined in these folders before use. All methods are located in the models folder and inherit from the RecommenderBase class in dgl_recommender_base.py

Look in datasets folder for the construction of datasets.

NOTICE

Section describing modifications to documents under the Apache License or for recognition. Both IGMC and NGCF are originally under the Apache License.

Models

IGMC

The IGMC model is a modified version of Based on https://github.com/LspongebobJH/dgl/blob/pr1815_igmc/examples/pytorch/igmc/ For download: https://github.com/LspongebobJH/dgl/archive/refs/heads/pr1815_igmc.zip

Specifically, it has been implemented with the RecommenderBase and uses a DGL collator for training.

KGAT

Heavily inspired by the work of jennyzhang0215 in https://github.com/jennyzhang0215/DGL-KGAT.

NGCF and LightGCN

Based on the example of NGCF in Based on https://github.com/dmlc/dgl/tree/master/examples/pytorch/NGCF modified to include LightGCN and using blocks.

DGL

Under models/shared/dgl_dataloader.py, you will find changes of the DGL EdgeCollator classes and EdgeDataloader to skip certain graph building steps.