Fairness-Aware Online Positive-Unlabeled Learning

Data Preparation

Install necessary packages.

pip install datasets
pip install gensim
python -m spacy download en

Run preprocessing code,

python chat_preprocessing.py
python wiki_preprocessing.py

For a pre-trained encoder, e.g. BERT, run BERT-version preprocessing,

python bert_chat_preprocessing.py
python bert_wiki_preprocessing.py

Training

Run

python main.py --model MODEL --dataset DATASET --r1 R1 --batch_size BATCH_SIZE --epoch EPOCH --pu_type PU_TYPE --lr LR --round ROUND --fairness FAIRNESS_CONSTRAINT --online --fair --lam_f LAMBDA_F --penalty --lam_penalty --LAM_PENALTY

Options

• MODEL: mlp and linear.
• DATASET: wiki, chat_toxicity
• R1: the ratio of unlabeled samples in positive instances.
• BATCH_SIZE: the batch size for offline learning. The batch size will be set as a full batch in online learning.
• EPOCH: the number of epochs for offline learning. The number of epochs will be set as 1 in online learning.
• PU_TYPE: upu and nnpu.
• LR: lr=0.01, 0.1 is recommended for offline learning. lr=0.01, 0.1, 1.0 is recommended for online learning.
• ROUND: The number of total rounds for online learning.
• ONLINE: command --online to run online learning. To execute offline learning, drop this command.
• FAIR: command --fair to use fairness loss. To execute the baseline, drop this command.
• FAIRNESS_CONSTRAINT: fairness constraint to compute a fairness loss. ddp and deo.
• LAMBDA_F: a hyperparameter to weigh the fairness loss.
• PENALTY: command penalty to use penalty term.
• LAM_PENALTY: a hyperparameter to weigh the penalty term.

Example

python main.py --dataset chat_toxicity --online --model linear --pu_type nnpu --gpu_id 1 --lr 1 --r1 0.5 --fair --fairness deo --lam_f 0.0001 --penalty --lam_penalty 1