Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation

This repository contains the Zenodo link to our released proprietary dataset as well as the Python code of our model from the paper Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation, accepted for publication in the proceedings of the 18th ACM Conference on Recommender Systems (RecSys 2024). The paper is available online on arXiv.

Environment

• python 3.9.13
• tensorflow 2.11.0
• tqdm 4.65.0
• numpy 1.24.2
• scipy 1.10.1
• pandas 1.5.3
• toolz 0.12.0

Deezer Dataset

The original anonymized version of our Deezer proprietary dataset (before filters applied in this work) can be freely downloaded from Zenodo. This dataset contains over 700 million time-stamped listening events collected from 3.4M anonymised users on Deezer streaming service, occurred between March and August 2022. It includes 50k anonymised songs, among the most popular ones on the service as well as their pre-trained embedding vectors, calculated by our internal model. All files are in parquet format which could be read by using pandas.read_parquet function.

General Architecture

Hyperparameters

Hyperparameters on each dataset are found in the corresponding configuration file in configs directory.

Experiments

Pretrained Embeddings

For LFM1B, if the pretrained embeddings are used ("pretrained" in configuration file is set to "item", otherwise "nopretrained"), we first need to generate pretrained embeddings as following:

1. Download data and put it into exp/data directory. For example exp/data/lfm1b
2. From recsys24-pisa directory, run python scripts in data_misc/lfm1b directory in sequence (from 1 -> 5)
3. The output pretrained embeddings will be found in exp/data/lfm1b

For Deezer dataset, the pretrained embeddings are calculated beforehand by our internal model and are provided alongside with user sessions.

Main scripts

To run the experiments, from the root directory in the terminal (recsys24-pisa), run the command ./script/run_pisa.sh

Some results

Repeat- vs. Non-repeat-aware

LFM1B

DEEZER

PISA vs. Other repeat-aware baselines

LFM1B

DEEZER

Repetition / Exploration Biases

Popularity Bias

Role of ACT-R components

Cite

Please cite our paper if you use this code in your own work:

@inproceedings{tran-recsys2024,
  title={Transformers Meet ACT-R: Repeat-Aware and Sequential Listening Session Recommendation},
  author={Tran, Viet-Anh and Salha-Galvan, Guillaume and Sguerra, Bruno and Hennequin, Romain},
  booktitle = {Proceedings of the 18th ACM Conference on Recommender Systems},
  year = {2024}
}