Practical Work in AI: Personalizing item relevance with psychology-based interest for repeat-aware music recommender systems

Pre-requisites

• miniconda installed
• GPU available
• CUDA 11.5 or higher
• Python 3.8 or higher installed

In order to get started with the code, follow these steps:

1. Clone GitHub repository

Clone this repo and initialize the submodules via this command:

git clone --recurse-submodules https://github.com/Laura-Legat/Practical-Work-AI

2. Create miniconda environment such that the libraries work

Via this command (executed from the root folder):

conda env create -f environment.yml

And activate:

conda activate pr_ai

3. Run the setup to install location-independent custom code

Via this command (executed from the root folder):

python setup.py install

4. Create dataset files

Via this command (executed from the root folder):

python preprocess.py -sl 50 -st 1

The -sl flag is the sequence length, -st the stride. Feel free to adapt these to your liking. The output files will be stores in the /data/ subdirectory.

Hyperparameter optimization

Perform hyperparameter optimization for any number of trials for either of the models with the optuna_paropt.py script. Please refer to the argparse documentation within the script for further information on the flags. The best parameters as well as trial information will be stored in the /optim/ subdirectory. For this step to work, please create ex2vec_search_space.csv and gru4rec_search_space.csv files beforehand. Search space, optuna visualization as well as best trained model visualization information will eb stored in the /results/ subdirectory. To modify the search spaces, modify the files in the /paramspaces/ subdirectory located in the GRU4Rec_Fork submodule folder.

Training

Train/finaltrain Ex2Vec by running the train.py script. Similarly, train GRU4Rec by running the run.py script located in the GRU4Rec_Fork submodule folder. Please refer to the argparse documentation within the script for further information on the flags. Resulting models will either be stored as checkpoints in their respective chckpts subdirectories (any checkpoint models during training), or in the models subdirectory (final model). For Ex2Vec, tensorboard information will be stored in the /runs/ subdirectory.

Combination ideas

Combination idea 1 and 3 can be run via their respetive scripts located in the root folder. Combination idea 2 can be run by adding the -ep flag when executing train.py to specify a GRU4Rec model path. Please refer to the argparse documentation within the script for further information on the flags.

Dataset

The dataset (new_release_stream.csv) is provided here.

References

• Official README of Ex2Vec repository
• Official README of GRU4Rec PyTorch implementation repository
• Official README of GRU4Rec Theano implementation repository

@inproceedings{sguerra2023ex2vec,
  title={Ex2Vec: Characterizing Users and Items from the Mere Exposure Effect},
  author={Sguerra, Bruno and Tran, Viet-Anh and Hennequin, Romain},
  booktitle = {Proceedings of the 17th ACM Conference on Recommender Systems},
  year = {2023}
}

Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk: Session-based Recommendations with Recurrent Neural Networks. arXiv preprint arXiv:1511.06939, 2015. https://arxiv.org/abs/1511.06939 Presented at the 4th International Conference on Learning Representations, ICLR 2016.

Balázs Hidasi, Alexandros Karatzoglou: Recurrent Neural Networks with Top-k Gains for Session-based Recommendations. arXiv preprint arXiv:1706.03847, 2017. https://arxiv.org/abs/1706.03847