/bert4rec_repro

Primary LanguagePython

Our fork used to reproduce Bert4Rec's results on Beauty, and provide better SASRec results by improving the loss function.

Preparation

  1. git clone the aprec repo.
git clone https://github.com/asash/bert4rec_repro.git

  1. follow its README.md to install relevant dependencies and test them.

  2. add our implement:

  • Replace the bce.py file in the losses/ folder with the bce.py file we provided.
  • Replace the negative_per_positive_target.py file in the recommenders/dnn_sequential_recommender/target_builders/ folder with the negative_per_positive_target.py file we provided.
  • Add our ml-1m-benchmark.py file at evaluation/configs/bert4rec_repro_paper/ folder.

Better SASRec result.

Table2

Run the command below at evaluation/ folder to get our experiment result of Table 3.

CHECK_COMMIT_STATUS=false sh run_n_experiments.sh configs/bert4rec_repro_paper/ml-1m_benchmark.py

Replicate Beauty result.

Note if you want to replicate the BERT4Rec results on the Beauty datasets can follow the step bellow:

  • Add ./beauty_s3.txt to data/bert4rec/ folder.
  • Change 'datasets/bert4rec_datsets.py' file.

tmp

  • Change 'datasets/datasets_register.py' file.

tmp

  • Change the DATASET at config file 'ml-1m-benchmark.py'
  • Run command CHECK_COMMIT_STATUS=false sh run_n_experiments.sh configs/bert4rec_repro_paper/ml-1m_benchmark.py