Revisiting IPS in Recommendation Models: Unveiling Its Impact on Model Performance

This repository is the official implementation of Revisiting IPS in Recommendation Models: Unveiling Its Impact on Model Performance with pytorch.

1. Preparation

• All package dependencies and their versions for reproducing our experimental environment are listed in requirements.txt file.
• The KuaiRec dataset is available for download via the provided Google Drive or the official release page (https://kuairec.com).

no_ips-0531
├──interaction/
│   └──data/
│      ├──KuaiRec/
│      │   ├──data/
│      │   │  ├──big_matrix.csv
│      │   │  ├──...
│      │   └──...
│      ├──yahoo_r3/
│      └──...
├──causality/
├──...

• Due to storage limitations, DunnHumby: Original, Personalized datasets need to be downloded at here.
• Data Organization for Experiments of Causality-based Recommendation:

no_ips-0531
├──causality/
│   └──data/
│      └──dunn_cat_mailer_10_10_1_1/
│         ├──original_rp0.40/
│         │  ├──data_test_000.csv
│         │  ├──...
│         │  ├──data_test_057.csv
│         │  ├──data_vali_000.csv
│         │  ├──...
│         │  └──data_vali_018.csv
│         └──rank_rp0.40_sf2.00_nr210/
│            ├──data_test_000.csv
│            ├──...
│            ├──data_test_056.csv
│            ├──data_vali_000.csv
│            ├──...
│            └──data_vali_021.csv
├──interaction/
├──...

• After dowloading the full data, please run merge_data.py.

2. Main Results

• The following files reproduce the main results and include all the hyperparameter settings.

report_table1.sh
report_table3.sh
report_table5.sh
report_table6
report_fig3.sh

3. Code Examples

• The following files provide brief code examples demonstrating how to execute the experimental pipeline.

ex_interaction.ipynb
ex_causality.ipynb