Warm-start framework

Data loaders and model training/evaluation pipelines written using Python 3.8 used in the MindReader paper published at CIKM 2020, https://doi.org/10.1145/3340531.3412759. You can read more about our dataset at https://mindreader.tech/dataset, and remember to cite our work:

@inproceedings{brams2020mindreader,
  title={MindReader: Recommendation over Knowledge Graph Entities with Explicit User Ratings},
  author={Brams, Anders H and Jakobsen, Anders L and Jendal, Theis E and Lissandrini, Matteo and Dolog, Peter and Hose, Katja},
  booktitle={Proceedings of the 29th ACM International Conference on Information \& Knowledge Management},
  pages={2975--2982},
  year={2020}
}

Get the most recent data

Run the ./data_loading/download_raw_data.py script to download the most recent MindReader data.

Generate datasets

Run the ./data_generation_entry.py script to generate data. Consult generate() to adjust whether or not to include top-popular items in the test set.

Running the experiments

Run the ./run.sh script for running all models in all experiments. Results are written to ./results/.

Running specific models and/or experiments

First, build the Docker image:

docker build -t mi911/runner .

When running the container, you have the following options:

1. --include [MODEL NAME LIST] for running only specific models (defaults to all models)
   • Model names: item-knn, user-knn, mf, svd, bpr, transe, transe-kg, transh, transh-kg, ppr-collab, ppr-kg, ppr-joint, random, top-pop, and cbf-item-knn.
2. --exclude [MODEL NAME LIST] for running all models except specific ones (defaults to none)
3. --experiments [EXPERIMENT NAME LIST] for the experiments to run
   • Experiment names (prefixed wtp- and ntp- for with and without top-popular items in the test set, respectively): all_movies, all_entities, substituting-3-4, substituting-2-4, and substituting-1-4.
4. --debug for printing debug-level logs to the terminal.

For example, if we want to run the experiment containing all movie ratings with top-popular items in the test set running only the SVD and BPR models, the following command will work:

docker run -d -v ${PWD}/.data:/app/data -v ${PWD}/results:/app/results mi911/runner --include bpr svd --experiments wtp-all_movies