Fairness Encoding in Logic Tensor Networks (FEL)

FEL is a debiasing framework for feature-based binary classification tasks.

General Description

• FEL integrates fairness and accuracy objectives within the Logic Tensor Networks (LTN) framework, where learning is guided by logical axioms rather than direct loss modifications.
• Through classification axioms, the model learns to correctly represent the relationship between features and target labels, ensuring predictive accuracy.
• Fairness axioms introduce constraints that promote statistical parity between sensitive groups, embedding fairness directly into the training process.
• Evaluation combines both perspectives:
  • Fairness is assessed using Statistical Parity Distance (SPD) and Disparate Impact (DI).
  • Performance is measured through overall classification accuracy.

Top Features

• Declarative representation of constraints: no need to modify the loss function
• Compositionality: can be generalised to different fairness definitions
• Actively control the level of fairness
• No need to transpose the definition into a metric

Setup Instructions

1. Prerequisites

1. Install Docker
2. Create a YAML config file (example: ./config.yaml)
3. Prepare a CSV dataset with headers

2. Build the Docker image

$ docker build -t my_env .

3. Run the container

$ docker run --gpus all -it -v $(pwd):/home/developer/ -w /home/developer my_env bash

4. Inside the container

Train a fair model:

$ python src/training.py --config ./config.yaml --dataset ./datasets/compas.csv

Run inference:

$ python src/inference.py --config ./config.yaml --dataset ./datasets/compas.csv --model ./models/kb.npz

Configuration Guide

The required YAML file need to specify the following parameters:

Data Configuration

• target_variable: Column name containing binary outcomes
• sensitive_feature: Column name of the protected attribute
• labels:
  • positive: The outcome value to balance fairly across groups
  • negative: The other possible outcome value
• protected_values:
  • privileged: Value(s) marking the privileged group
  • unprivileged: Value(s) marking the unprivileged group

Model Configuration

• hidden_layer_sizes: Neural network architecture [list of integers]
• implies: LTN fuzzy implication operator:
  • Options: KleeneDienes, Godel, Reichenbach, Goguen, Luk
• p_mean: Quantifier aggregation parameter:
  • Options: [1,3,5,7]
• aggregator_deviation: Deviation control in aggregator
• fairness_weight: Weight of fairness axioms wrt classification axioms.
  • Options: [1...3]: A value of 1 gives fairness and classification axioms equal weight, while 3 assigns 3x times importance to fairness satisfiability.

Training Configuration

• epochs: Number of training iterations

Understanding "Positive" and "Negative" Labels

The choice of positive/negative labels is crucial:

• Positive: The outcome you want to balance fairly across groups
  • Example: loan_approved=1 means "approved"
  • Fairness metrics measure this outcome's distribution
• Negative: The alternative outcome
  • Not directly balanced by fairness constraints

Examples:

• Loan approval: positive = 1 (approved), negative = 0 (denied) → approval rates should be similar across groups.

• Disease detection: positive = “has_disease”, negative = “no_disease” → detection rates should be comparable across groups.

License

FEL is open-source and released under the Apache-2.0 License.

Cite this Work

Greco, G., Alberici, F., Palmonari, M., & Cosentini, A. (2023). Declarative Encoding of Fairness in Logic Tensor Networks. Frontiers in Artificial Intelligence and Applications, 372, 908–915. IOS Press BV. https://boa.unimib.it/bitstream/10281/462523/1/Greco-2023-ECAI-VoR.pdf