Code for CIKM-2024 Full Paper:
TEXT CAN BE FAIR: Mitigating Popularity Bias with PLMs by Learning Relative Preference

STEP 0: Prepare the environment

• conda create -n staple python=3.8
• conda activate staple
• while read requirement; do pip install "$requirement"; done < requirements.txt (in Linux)
• install torch_scatter through .whl file from https://pytorch-geometric.com/whl/, such as:
  • wget https://data.pyg.org/whl/torch-1.13.0%2Bcu117/torch_scatter-2.0.9-cp38-cp38-linux_x86_64.whl
  • pip install torch_scatter-2.0.9-cp38-cp38-linux_x86_64.whl

STEP 1: Prepare the dataset.

Please keep the following folder structure, download the dataset from the official URL in the paper, and download the pre-trained model from the Huggingface:

-staple.py
-run_staple.py
...
-dataset
  -Metadata
    -meta_Office_Products.json.gz
  -Ratings
    -Office_Products.csv
-plm
  -opt-125m

STEP 2: Preprocess the dataset.

• python preprocess_amazon.py --dataset=Office

STEP 3: Pre-training the ranker model

• For multi GPUs:
  • nohup python run_staple.py --dataset=Office --num_gpus=6 --batch_size=16 --distributed --multiGPU --valid_ratio=0.1 --train_stage=1 > Office_stage1.log 2>&1 &
• For single GPU:
  • nohup python run_staple.py --dataset=Office --batch_size=16 --valid_ratio=0.1 --train_stage=1 > Office_stage1.log 2>&1 &

STEP 4: Fine-tuning the ranker model

• For multi GPUs:
  • nohup python run_staple.py --dataset=Office --num_gpus=6 --batch_size=16 --distributed --multiGPU --valid_ratio=0.1 --train_stage=2 > Office_stage2.log 2>&1 &
• For single GPU:
  • nohup python run_staple.py --dataset=Office --batch_size=16 --valid_ratio=0.1 --train_stage=2 > Office_stage2.log 2>&1 &

STEP 5: Self-distillation

• For multi GPUs:
  • nohup python run_staple.py --dataset=Office --num_gpus=6 --batch_size=16 --distributed --multiGPU --valid_ratio=0.1 --train_stage=3 > Office_stage3.log 2>&1 &
• For single GPU:
  • nohup python run_staple.py --dataset=Office --batch_size=16 --valid_ratio=0.1 --train_stage=3 > Office_stage3.log 2>&1 &

Notes:

• In our experiments, we use 6 gpus and the batch size on each gpu is 16. Thus the total batch size is 96.
• The batch size on each gpu is important in stage 3 because the value of sampled cross-entropy loss is related to the negatives, and we sample (batch_size * 10) negative items for each batch. Changing the batch size may not get the ideal output.
• We have tested the codes for a quick start and reproducing the results reported in the paper. If you have any questions and find any bug, please let us know in the review, and we will fix the bug in the next version
• If you want to use OPT-125M directly from Hugging Face instead of downloading it locally, simply set the root_path to an empty string ("").