TLDR: TABFAIRGDT is a new method for creating fair synthetic tabular data using autoregressive decision trees. It addresses bias in machine learning models by employing a soft leaf resampling technique to ensure fairness while preserving data utility and quality. The method is non-parametric, requires no pre-processing, and is highly efficient, achieving a 72% average speedup over existing deep generative models. It can generate fair data for medium-sized datasets in one second on a standard CPU and avoids creating unrealistic ‘out-of-distribution’ samples, making it suitable for critical real-world applications.
Machine learning models are increasingly used in critical areas like finance and healthcare, but they often carry biases from their training data. These biases can lead to unfair outcomes for certain demographic groups based on sensitive attributes such as race, gender, or age. Addressing this issue is crucial, and one promising approach is to mitigate bias at the data level by generating fair synthetic data.
Many existing methods for generating fair synthetic data rely on complex deep learning architectures, which can be computationally intensive. However, new research suggests that simpler models can be highly effective for tabular data, which is the most common data type in real-world applications.
Introducing TABFAIRGDT: A New Approach to Fair Data Generation
A novel method called TABFAIRGDT (Tabular Fair Generative Decision Trees) has been introduced to generate fair synthetic tabular data using autoregressive decision trees. This approach stands out because it is non-parametric, meaning it doesn’t make assumptions about the underlying data distributions, making it ideal for mixed tabular data types. It also requires no data pre-processing, simplifying its application.
The core of TABFAIRGDT involves an autoregressive framework where features are generated sequentially using decision trees. To ensure fairness, the method incorporates a unique ‘soft leaf resampling’ technique. This technique adjusts the decision tree outputs to reduce bias, particularly when generating the target attribute, while carefully preserving the data’s predictive performance. This soft adjustment allows for a smoother balance between achieving fairness and maintaining the utility of the data.
Key Advantages and Performance
TABFAIRGDT has demonstrated significant advantages over existing state-of-the-art deep generative models. It achieves a better fairness-utility trade-off for downstream tasks and produces higher quality synthetic data. On average, it improves fairness by approximately 50% while maintaining a high predictive score.
One of the most remarkable aspects of TABFAIRGDT is its efficiency. It is lightweight, CPU-compatible, and incredibly fast. The method achieves an average speedup of 72% compared to the fastest baseline, and it can generate fair synthetic data for medium-sized datasets (10 features, 10,000 samples) in just one second on a standard CPU. This makes it an ideal solution for real-world applications where speed and accessibility are crucial.
Furthermore, TABFAIRGDT avoids generating ‘out-of-distribution’ samples that are specific to sensitive attributes. This is a critical feature, especially in sensitive domains like healthcare, where generating unrealistic feature combinations (e.g., a male being pregnant) could be problematic. By generating the sensitive attribute only after all other features are synthesized and enforcing fairness solely for the target generation, TABFAIRGDT respects real-world constraints and maintains data fidelity.
The research also found that the order in which features are generated does not significantly impact the results, adding to the method’s robustness and ease of use.
Also Read:
- Causal Representation Learning: Leveraging What We Can See
- CoUn: A Smarter Way for AI Models to Forget Specific Data
Looking Ahead
The introduction of TABFAIRGDT marks a significant step forward in generating fair synthetic tabular data. Its combination of effectiveness, efficiency, and non-parametric nature makes it a powerful tool for mitigating bias in machine learning models. Future work aims to expand its capabilities to support additional fairness definitions, handle continuous sensitive attributes, and explore the impact of various hyperparameters. For more technical details, you can refer to the full research paper here.
