Fake News: 0 Real News: 1
Left Leaning: Privileged Group: 0 Right Leaning: Unprivileged Group: 1
TN: predicted = 0 (fake), actual = 0(fake)
FN: predicted = 0(fake), actual = 1(real)
FP: predicted = 1 (real), actual = 0 (fake)
FN: predicted = 1 (real), actual = 1(real)
SPD = 0 to be fair EOD = 0 to be fair DIR = 1 to be fair AOD = 0 to be fair
Debiasing using ROC Threshold: lies between 0 and 1 Margin: lies between 0 and min(threshold, 1-threshold )
# steps to be done only the first time
pip install .
cd src
# To Train a different variant of the model or different hyperparameters, etc.
python train.py
# utility functions for metrics and debiasing
utils.py
# To test bias metrics - sklearn + aif360
python metrics.py
# To debias (posthoc debiasing using Reject Option Classifier)
python debias_input_prepare.py
python debias_roc.py
# To perform analysis on BERT using captum
python bert_analysis.py
# To extract important tokens using SHAP
python out_of_shap.py
# To extract important tokens using LIME
python in_the_lime.py
# To extract important tokens using Integrated Gradience
python int_grad.py
# to extract important phrases using SHAP and LIME
python phrase_level_shap.py
python phrase_level_lime.py
# To extract important sentences using SHAP
python sentence_level_shap.py
# To perform data injection attacks
python freq_hamla.py, python salience_hamla.py
python injection_attack_test.py
python attack_metrics.py