FairFinGAN: Fairness-aware Synthetic Financial Data Generation

The paper proposes FairFinGAN, a WGAN-based framework that integrates fairness constraints via a classifier to generate synthetic financial data that effectively mitigates bias against protected attributes while maintaining high utility for downstream predictive tasks.

The Gist

Imagine you are a bank manager trying to decide who gets a loan. You have a massive pile of historical application forms (the data) to help you make these decisions. But here's the problem: that pile of forms is messy. It contains old prejudices. Maybe the bank used to reject people from a certain neighborhood or age group unfairly, and those old mistakes are baked into the data. If you train a computer robot to make decisions based on this "dirty" data, the robot will just learn to be unfair too.

Now, imagine you want to test new ideas or train better robots, but you can't share the real customer data because of privacy laws (like GDPR). You need a fake version of the data that looks and acts exactly like the real thing, but without the secrets. This is called Synthetic Data.

The problem? If you just use a standard AI to make this fake data, it might accidentally copy the unfair biases from the real data, or even make them worse. It's like photocopying a biased document; the copy is just as biased as the original.

Enter: FairFinGAN (The "Fairness Filter" Chef)

This paper introduces a new tool called FairFinGAN. Think of it as a super-smart chef who doesn't just cook a meal that looks like the original recipe, but also ensures the meal is fair to everyone eating it.

Here is how it works, broken down into simple steps:

1. The Two-Phase Cooking Process

Most AI data generators are like a chef who just tries to mimic the taste of a dish perfectly. FairFinGAN does this in two distinct phases:

• Phase 1: The "Taste Test" (Making it Real)
  The AI (the Generator) tries to create fake financial records that are indistinguishable from real ones. It's like a forger trying to make a fake bill that looks exactly like a real one. A "Critic" (another AI) acts as a strict food critic, tasting the fake data and saying, "This doesn't taste like the real thing!" The Generator keeps trying until the Critic can't tell the difference.

  • Goal: Make the data look real so it's useful for testing.

• Phase 2: The "Fairness Check" (The Secret Sauce)
  This is the magic part. Once the data looks real, a third AI (a Classifier) steps in. This AI is trained to predict outcomes (like "Will this person pay back the loan?"). But here's the twist: the Generator is now being punished if the Classifier treats different groups of people (like men vs. women, or young vs. old) differently.

  • The Analogy: Imagine the Generator is a teacher creating practice exams. In Phase 1, they make sure the questions are hard and realistic. In Phase 2, a "Fairness Inspector" checks the exams. If the Inspector sees that the questions accidentally make it harder for students from Group A than Group B, the Teacher has to rewrite the questions. The Generator learns to tweak the data until the "Fairness Inspector" is happy.

2. The "Fairness" Metrics

The paper focuses on two main ways to measure fairness:

• Statistical Parity: This is like saying, "If 50% of Group A gets a loan, 50% of Group B should also get a loan, regardless of their actual credit score." It ensures equal opportunity at the surface level.
• Equalized Odds: This is a bit more nuanced. It says, "If a person is actually a good borrower, they should have the same chance of getting a loan, no matter which group they belong to." It ensures the AI isn't making mistakes more often for one group than another.

FairFinGAN can be tuned to prioritize either of these rules.

3. The Results: A Balanced Diet

The researchers tested this on five real-world financial datasets (like credit card defaults and credit scoring). They compared their "Fair Chef" (FairFinGAN) against other popular AI data generators.

• The Old Way (Standard AI): Often created data that was either very realistic but still unfair, or very fair but useless (because it didn't look like real data anymore).
• The FairFinGAN Way: It found the "Goldilocks" zone. It created data that was:
  1. Realistic enough to train good predictive models (high utility).
  2. Fair enough to reduce discrimination against protected groups (like age, gender, or race).

Why Does This Matter?

In the real world, banks and financial institutions are under pressure to be fair and to protect customer privacy.

• Privacy: They can share this "Fair Synthetic Data" with researchers without leaking real customer secrets.
• Fairness: They can use this data to train their loan-approval algorithms to be less biased, helping to break the cycle of historical discrimination.

The Bottom Line

FairFinGAN is like a smart editor for financial data. It takes a messy, biased, and private pile of information, and rewrites it into a clean, fair, and realistic story that anyone can use to build better, more equitable financial systems. It proves that you don't have to choose between data that is useful and data that is fair; you can have both.

Technical Summary

1. Problem Statement

Financial datasets are critical for automated decision-making systems (e.g., credit scoring, loan approvals) but often suffer from inherent biases related to protected attributes such as gender, race, and age. These biases can lead to unfair outcomes and discriminatory practices.

• Data Scarcity & Privacy: Real financial data is difficult to share due to privacy regulations and ownership constraints, necessitating the use of synthetic data.
• Bias Amplification: Existing Generative Adversarial Network (GAN) methods for synthesizing tabular data often reproduce or even amplify the biases present in the original training data.
• The Gap: There is a need for a framework that generates synthetic financial data which is both high-fidelity (preserving statistical utility for downstream tasks) and fair (mitigating bias at the dataset level before any classifier is trained).

2. Methodology: FairFinGAN

The authors propose FairFinGAN, a framework based on Wasserstein GANs (WGAN) designed to generate fair synthetic tabular data. The core innovation is a two-phase training strategy that integrates fairness constraints directly into the generator's optimization process.

A. Architecture

The model consists of three main neural network components:

1. Generator ($G$): A Multi-Layer Perceptron (MLP) that takes a latent noise vector $z$ and outputs synthetic samples $(x', y', s')$, where $x'$ are features, $y'$ is the class label, and $s'$ is the protected attribute. It uses Gumbel-Softmax to handle categorical data.
2. Critic ($C$): A WGAN critic that distinguishes between real and synthetic data to ensure the generated distribution matches the real data distribution.
3. Classifier ($H$): An MLP trained first on the original real dataset. This classifier is frozen during the second phase of training and used to evaluate the fairness of the synthetic data.

B. Two-Phase Training Process

The training occurs in two distinct phases:

• Phase 1: Data Synthesis (Accuracy Focus)

  • The Generator and Critic engage in standard adversarial training.
  • The goal is to generate synthetic data that closely mimics the statistical properties of the original dataset.
  • The Generator is updated to minimize the distance between real and synthetic distributions (Wasserstein distance).

• Phase 2: Fairness Optimization (Fairness Focus)

  • The Generator is further fine-tuned to minimize a fairness penalty.
  • The frozen Classifier $H$ (trained on real data) predicts labels for the synthetic data generated by $G$.
  • A fairness metric is calculated based on these predictions and the protected attributes of the synthetic data.
  • This metric is added to the Generator's loss function with a weighting parameter $\lambda_{fair}$.
  • The Generator is updated to produce data where the classifier's predictions are independent of the protected attribute.

C. Fairness Constraints

The framework supports two specific fairness definitions, implemented via the loss function in Phase 2:

1. Statistical Parity (SP): Ensures the probability of a positive prediction is equal across different groups of the protected attribute.
   • Implementation: FairFinGAN-SP.
2. Equalized Odds (EOd): Ensures that both the true positive rate and false positive rate are equal across groups.
   • Implementation: FairFinGAN-EOd.

3. Key Contributions

1. Novel Framework: Introduction of FairFinGAN, a WGAN-based framework specifically tailored for fairness-aware synthetic financial data generation.
2. Training Strategy: A unique two-phase approach that decouples data fidelity (Phase 1) from fairness optimization (Phase 2), using a pre-trained classifier to enforce fairness constraints without retraining the classifier itself.
3. Comprehensive Evaluation: Extensive experiments on five real-world financial datasets (Adult, Credit Card, Credit Scoring, Dutch Census, German Credit) comparing the method against state-of-the-art baselines (CTGAN, TabFairGAN).
4. Balanced Performance: Demonstration that the method achieves superior fairness metrics while maintaining high data utility (predictive accuracy) for downstream tasks.

4. Experimental Results

The authors evaluated the generated data using four standard classifiers (Logistic Regression, Decision Tree, k-NN, MLP) and measured both Accuracy/Balance Accuracy and 7 Fairness Metrics (SP, EO, EOd, PP, PE, TE, ABROCA).

• Fairness Performance:

  • FairFinGAN consistently achieved the best or second-best fairness scores across most datasets and protected attributes.
  • FairFinGAN-EOd generally provided better results for Equalized Odds and Equal Opportunity, while FairFinGAN-SP excelled in Statistical Parity.
  • Compared to TabFairGAN, FairFinGAN offered a better trade-off; TabFairGAN often achieved extreme fairness (near-zero disparity) but at the cost of significantly degraded predictive utility (low accuracy).
  • Compared to CTGAN, FairFinGAN significantly reduced bias without the severe accuracy drops sometimes seen in fairness-focused baselines.

• Data Utility:

  • Models trained on FairFinGAN synthetic data maintained competitive accuracy levels compared to models trained on real data.
  • In several cases (e.g., Credit Card dataset with "Sex" attribute), FairFinGAN-SP achieved the highest accuracy among all methods while maintaining strong fairness.

• Robustness: The method proved effective across different protected attributes (Gender, Race, Age, Sex) and varying levels of class imbalance.

5. Significance and Future Work

• Significance:

  • Bias Mitigation at the Source: By addressing fairness at the data generation level rather than the classification level, FairFinGAN provides a robust solution for creating unbiased datasets for research and development.
  • Regulatory Compliance: The tool aids financial institutions in aligning automated decision-making with regulatory requirements (e.g., GDPR, fair lending laws) by providing synthetic data that is legally safer to use and share.
  • Privacy Preservation: It enables the sharing of high-quality financial data without exposing sensitive real-world records.

• Future Directions:

  • Extending the framework to handle multiple protected attributes simultaneously.
  • Applying the method to other critical domains like healthcare and education.
  • Integrating differential privacy to further enhance data security.
  • Exploring more advanced fairness metrics beyond the standard statistical definitions.

In conclusion, FairFinGAN represents a significant step forward in ethical AI for finance, successfully bridging the gap between the need for high-quality synthetic data and the imperative of algorithmic fairness.