Learning Controllable Fair Representations (1812.04218v3)

Abstract: Learning data representations that are transferable and are fair with respect to certain protected attributes is crucial to reducing unfair decisions while preserving the utility of the data. We propose an information-theoretically motivated objective for learning maximally expressive representations subject to fairness constraints. We demonstrate that a range of existing approaches optimize approximations to the Lagrangian dual of our objective. In contrast to these existing approaches, our objective allows the user to control the fairness of the representations by specifying limits on unfairness. Exploiting duality, we introduce a method that optimizes the model parameters as well as the expressiveness-fairness trade-off. Empirical evidence suggests that our proposed method can balance the trade-off between multiple notions of fairness and achieves higher expressiveness at a lower computational cost.

Learning Controllable Fair Representations: An Information-Theoretic Approach

The paper presents an innovative approach to learning data representations that are both expressive and fair concerning protected attributes using an information-theoretically motivated framework. The focus is on developing a method that allows the user to control the level of fairness in the learned representations while maximizing their expressiveness. This dual objective ensures that the representations can be used effectively for multiple tasks without perpetuating unfair biases associated with sensitive attributes.

Key Contributions

1. Information-Theoretic Framework: The paper introduces a constrained optimization framework grounded in information theory. The objective is to maximize the expressiveness of data representations while adhering to predefined fairness constraints. This is accomplished by utilizing mutual information to evaluate both expressiveness and fairness.
2. Expressiveness-Fairness Trade-Off: The paper addresses the inherent trade-off between expressiveness and fairness in data representation. By defining mutual information objectives for several dominant notions of fairness—including demographic parity, equalized odds, and equalized opportunity—the authors provide a method that allows a user to specify fairness constraints directly.
3. Duality and Optimization: Exploiting concepts from the Lagrangian duality, the proposed method optimizes both the model parameters and the expressiveness-fairness trade-off simultaneously. As a result, the framework provides direct user control over the fairness of representations through constraints that can be interpreted without domain expertise.
4. Unified Framework: This work serves as a unifying framework for existing methodologies in fair representation learning. It demonstrates that many previous approaches can be viewed as optimizing approximations of the Lagrangian dual of the proposed objective with fixed Lagrange multipliers.

Empirical Findings

The empirical evaluations in the paper suggest that the proposed controllable method can effectively balance multiple notions of fairness while maintaining high levels of expressiveness at reduced computational costs. Various experimental setups confirm that the learned representations align well with conventional definitions of fairness, such as demographic parity, but do so with the added advantage of direct user control over fairness levels.

Implications and Future Directions

The practical implications of this research are significant for applications where data needs to be released for various unspecified downstream tasks while ensuring fairness. By facilitating user-defined control over fairness parameters, the proposed framework enhances fairness-aware decision-making systems across industries like finance, education, and criminal justice.

Theoretically, this work paves the way for further exploration of mutual information as a measure in other complex fairness definitions and extends the application to non-binary sensitive attributes. Future research directions could investigate improving the tractability of the bounds applied or adopting more sophisticated distributions to enhance representation learning outcomes. Additionally, alternative training methods for the adversarial setup or more advanced model configurations may further improve the balance between fairness and expressiveness.