SentiFMRecSys

This repository includes the code and data associated to our UMAP '24 paper Integrating sentiment features in factorization machines: Experiments on music recommender systems. Dataset and pretrained models have been uploaded separately due to size constraints: dataset and pretrained models.

Steps

1. Set up a Conda environment with all the necessary libraries:

   • Create environment: conda env create -f environment.yml
   • Activate environment: conda activate lastfm_sentiment_umap_24
   • If CUDA (GPU) isn't available at first, consider rebooting the system

2. Train and evaluate models with the preprocessed dataset in dataset/lastfm_recbole-dataset.pth:

   • python3 recbole_run.py --dataset lastfm_recbole --config_files config/generic.yaml -m [Model]
   • Add --save option to save the model in "saved/" as .pth file

3. Train and evaluate models by preprocessing the dataset from scratch (takes about ~8GB RAM):

   • python3 recbole_run.py --dataset lastfm_recbole --config_files config/generic_unprocessed.yaml -m [Model]

4. Evaluate saved models:

   • python3 recbole_run.py --dataset lastfm_recbole --config_files config/generic.yaml --evaluate_model saved/[Saved Model].pth
   • Add --evaluation_mode [full | uniN | popN] to specify evaluation method, default is uni100
   • Some pretrained models are included, trained with the following embeddings:
     • v (valence)
     • a (arousal)
     • d (dominance)
     • stsc (sentiment ratio)

Citation

If you use our source code, dataset, or experiments for your research or development, please cite the following paper:

@inproceedings{wang2024umap,
  title={Integrating sentiment features in factorization machines: Experiments on music recommender systems},
  author={Javier Wang and Alejandro Bellogín and Iván Cantador},
  booktitle={{UMAP} '24: 32nd {ACM} Conference on User Modeling, Adaptation and Personalization, Cagliari, Italy, July 1 - 4, 2024},
  year={2024}
}