This repository provides the full implementation with released models. Requires python 3.7 and TensorFlow 2.0 (see requirements.txt).

Can we mitigate the unintended bias of cyberbullying detection models by guiding the model training with fairness constraints?

# list group-specific FNRs/FPRs
fnrs = []
fprs = []
constraints = []
for iden in range(cf.num_identity_groups):
    context_group_subset = context_group.subset(lambda kk=iden: group_tensor[:, kk] > 0)
    fnrs.append(tfco.false_negative_rate(context_group_subset))
    fprs.append(tfco.false_positive_rate(context_group_subset))

# define lower and upper bound constraints (see equation 3 in paper)
constraints.append(tfco.upper_bound(fnrs) - tfco.false_negative_rate(context) <= allowed_fnr_deviation)
constraints.append(tfco.upper_bound(fprs) - tfco.false_positive_rate(context) <= allowed_fpr_deviation)
constraints.append(tfco.false_negative_rate(context) - tfco.lower_bound(fnrs) <= allowed_fnr_deviation)
constraints.append(tfco.false_positive_rate(context) - tfco.lower_bound(fprs) <= allowed_fpr_deviation)

# define problem, optimizer and variables to optimize
problem = tfco.RateMinimizationProblem(objective, constraints)
optimizer = tfco.ProxyLagrangianOptimizerV2(
    optimizer=tf.keras.optimizers.Adam(learning_rate),
    constraint_optimizer=tf.keras.optimizers.Adam(learning_rate),
    num_constraints=problem.num_constraints)
var_list = (constrained_model.trainable_weights + problem.trainable_variables + optimizer.trainable_variables())

See directory_info in the data directory for the expected directory structure.

-Jigsaw	Link
-Twitter	Link
-Wiki	Link
-Gab	Link

Download released models to models directory. See directory_info in the model directory for the expected directory structure.

See source directory.

Run the corresponding notebook (e.g. gab_experiment.ipynb) for each experiment in the source directory for reproducing the full results from scratch. Note that the algorithms are non-determinisitic due to random weight initialization of the models.

Relevant configurations are defined in configs.py, e.g.:

--batch_size	128
--epochs	75
--gab_allowed_fnr_deviation
	0.10
--gab_allowed_fpr_deviation
	0.15
--random_state	42

source directory tree:

├── compare_models.ipynb
├── configs.py
├── embeddings.py
├── evaluation.py
├── gab_experiment.ipynb
├── jigsaw_experiment.ipynb
├── metrics.py
├── model.py
├── plot.py
├── train.py
├── twitter_experiment.ipynb
├── utils.py
└── wiki_experiment.ipynb

@article{gencoglu2021cyberbullying,
  title={Cyberbullying Detection with Fairness Constraints},
  author={Gencoglu, Oguzhan},
  journal={IEEE Internet Computing},
  doi={10.1109/MIC.2020.3032461},
  pages={20--29},
  volume={25},
  number={1},
  publisher={IEEE Computer Society},
  year={2021}
}

Or

Gencoglu, Oguzhan. "Cyberbullying Detection with Fairness Constraints." IEEE Internet Computing (2020) vol. 25, no. 1, pp. 20-29. doi:10.1109/MIC.2020.3032461