Codebase for the paper Improving Minimax Group Fairness in Sequential Recommendation

Structure

The source code is in the src directory and the data in the data directory.

1. The implementation of the sequential recommender and training methods is in src/recommenders.

   • src/recommenders/SASRec.py contains the SASRec recommender, and src/recommenders/utils/sequencer.py contains the sequencer used for padding/truncation of each user's item sequence.
   • The distributionally robust methods and baselines in the paper are available in src/recommenders/utils/loss.py.
   • src/recommenders/utils/metric_computer.py and src/recommenders/utils/checkpointer.py are used for calculating metrics and checkpointing the best model, respectively.

2. src/training/train_sasrec_dp and src/sagemaker/sagemaker_training.py are used to launch jobs locally and on AWS Sagemaker.

3. Finally, the improvements compared to standard training are in src/improvements. The RR_improvement.ipynb and ML1m_improvement.ipynb notebooks contain the analysis of the RetailRocket and Movielens1M experiments.

Datasets and Processing

We use Movielens1M and RetailRocket, which are popular open datasets for movies and e-commerce.

1. Download the processed ML1m from https://github.com/FeiSun/BERT4Rec/blob/master/data/ml-1m.txt and place it in data/raw/ml1m. Download the events.csv from https://www.kaggle.com/datasets/retailrocket/ecommerce-dataset and place it in data/raw/retailrocket.
2. We have already preprocessed the two datasets by running src/preprocess/preprocess_movielens1m.py and src/preprocess/preprocess_retailrocket.py, respectively. The processed data is in data/processed. We have also added user groups by running src/preprocess/addgroups_dsplit.py.

Usage

Note that the PYTHONPATH for your conda environment should be set to the base of this repository.

Local Run

We first suggest running python src/training/train_sasrec_dp.py, this script has default values for the training parameters.

The key parameters we change for the experiments:

• The loss type is configurable to one of: ["joint_dro", "erm", "cb", "cb_log", "s_dro", "group_dro", "ipw", "ipw_log"].
• joint-dro-alpha is the alpha level for CVaR DRO.
• The gdro-stepsize is the stepsize for the ascent step in the group and streaming dro loss.
• stream-lr is the streaming learning rate for SDRO.
• groups selects which user group and size we use in the experiment:
  • [popdsplit_balanced, popdsplit_0.2_0.6_0.2, popdsplit_0.1_0.8_0.1] are G_pop33, G_pop2060, and G_pop1080 from the paper.
  • [seqdsplit_balanced, seqdsplit_0.2_0.6_0.2, seqdsplit_0.1_0.8_0.1] are G_seq33, G_seq2060, and G_seq1080 from the paper.
  • popseq_bal
• subgroup-for-loss selects which subgroup to use for loss computation for GDRO, SDRO, and IPW when users belong to both popularity and sequence length groups. 0 uses popularity-based groups, 1 uses sequence length groups.

Launching Jobs on Sagemaker

Launching jobs on Sagemaker is straightforward via python src/sagemaker/sagemaker_training.py <config path> with the appropriate config path. The configs for all the jobs in our experiments are available in src/sagemaker/training_configs, for e.g. the command:

python src/sagemaker/sagemaker_training.py 'src/sagemaker/training_configs/RR/sasrec/popseq/rr-sasrec-jdro-popseq.json'

Launches CVaR DRO (also called jointDRO) training jobs with 10 different alpha values, on the RetailRocket dataset for users belonging to intersecting groups (popularity and sequence length-based groups).

---