A fork from Deep_CBRS_Amar_Revisited which follows the work from Deep_CBRS_Amar.
In this project, we experimented and evaluated different Graph Neural Network models to assess their performances in a recommendation task:
- How well GNNs perform when used as collaborative features extractors in contrast to using precomputed embeddings obtained by relying on Knowledge Graph Embeddings (KGEs) literature?
- How can we effectively integrate GNNs in a hybrid recommender system which leverages both collaborative and content-based features, i.e. also including both textual content and properties of items?
For further details, refer to the documentation
This repo requires at least Python 3.8
pip install -r requirements.txt
In this work, we used MovieLens-1M, preprocessed files can be obtained with DVC:
dvc pull dataset
However, for every dataset you'll need:
- CSV / TSV of ratings for train and test in the form
(user, item, liked / not liked)
For Hybrid architectures:
- JSON embeddings for content based features, respectively for user and item in the form:
[{"ID_OpenKE": <id_open_ke>, "profile_embedding": <embedding_in_list_form>}][{"ID_OpenKE": <id_open_ke>, "embedding": <embedding_in_list_form>}]
For knowledge-aware architectures:
- Triples CSV / TSV in the form
(item, entity, relation)
For knowledge-aware architectures:
- Triples CSV / TSV in the form
(item, entity, relation)
Run an experiment with experiment.py.
By default, config.yaml and experiments.yaml will be used for configuration.
python src/experiment.py
The following parameters can be specified as well:
- -c / --config: config input file, manages input parameters of the experiments;
- -e / --experiments: grid search experiment file, performs a grid search by overriding specified parameters in it with the ones in the config file;
- --exp_name: experiment name used in MLFlow that will encapsulate the runs.
For example, the following command runs a grid search on the basic recommender system architecture with Graph Neural Networks:
python src/experiment.py --exp_name myexp -e econfigs/basic-gnn.yaml
NOTE: you need to pull the dataset to run our experiments (↑) and the BERT embeddings for the Hybrid experiments:
dvc pull embeddings
Refer to the econfigs readme for an explanation for every grid configuration of runs available.
If you want to see our results, download them with DVC (around 40 GiB)
dvc pull mlruns
User MLFlow to compare experiments with its UI:
mlflow ui
Each run has its artifact folder where you can find the trained weights, the top 5/10 calculated predictions for the test set, logs and config.yaml to reproduce that exact run.