Themis-ml: A Fairness-aware Machine Learning Interface for End-to-end Discrimination Discovery and Mitigation

Abstract: As more industries integrate machine learning into socially sensitive decision processes like hiring, loan-approval, and parole-granting, we are at risk of perpetuating historical and contemporary socioeconomic disparities. This is a critical problem because on the one hand, organizations who use but do not understand the discriminatory potential of such systems will facilitate the widening of social disparities under the assumption that algorithms are categorically objective. On the other hand, the responsible use of machine learning can help us measure, understand, and mitigate the implicit historical biases in socially sensitive data by expressing implicit decision-making mental models in terms of explicit statistical models. In this paper we specify, implement, and evaluate a "fairness-aware" machine learning interface called themis-ml, which is intended for use by individual data scientists and engineers, academic research teams, or larger product teams who use machine learning in production systems.

• The paper introduces themis-ml, a fairness-aware machine learning interface and library with an API framework to facilitate discrimination discovery and mitigation in ML pipelines.
• Themis-ml provides techniques for preprocessing, model training (like Relabelling, Reweighting, PRR), and scoring metrics (like MD, NMD) to assess and reduce potentially discriminatory predictions.
• Evaluation on the German Credit Dataset demonstrates themis-ml's utility in measuring and mitigating bias, highlighting the practical fairness-utility tradeoff in real-world applications.

Analyzing Themis-ml: Interface for Fairness in Machine Learning

The paper introduces themis-ml, a fairness-aware machine learning interface designed to address discrimination in decision support systems (DSS). These systems, which integrate machine learning into socially sensitive processes such as hiring and loan approvals, carry the risk of perpetuating historical biases and socio-economic disparities. The research seeks to bridge the gap between machine learning’s transformative potential and its ability to propagate bias, by providing an interface that facilitates discrimination discovery and mitigation.

Key Contributions

1. API Propositions: The paper proposes an application programming interface (API) framework for Fairness-aware Machine Learning Interfaces (FMLI). Leveraging this, developers can integrate fairness-aware methods within their existing ML applications, particularly those employing binary classifiers.
2. Themis-ml Implementation: Introduced in the paper is themis-ml, an FMLI-compliant library centered around fairness-aware techniques. Utilizing this library, one can employ discrimination discovery methods within ML pipelines, thereby enabling the measurement and potential mitigation of biases inherent in historical data.
3. Evaluation and Potential Impact: Themis-ml is evaluated against the German Credit Dataset, emphasizing its capability to measure potentially discriminatory predictions (PD) and to mitigate PD using fairness-aware methods. This evaluation demonstrates the library's practical utility in real-world scenarios, particularly those involving social applications.

Examination of Bias and Discrimination

The paper delineates algorithmic bias as the tendency of mathematical rules within ML systems to favor certain attributes over others. This bias can lead to direct or indirect discrimination, known in legal contexts as disparate treatment and disparate impact. Disparate treatment can be mitigated by excluding attributes correlated with sensitive classes from training data. However, this approach does not address disparate impact, which stems from complex historical data processes.

Fairness-aware Interfaces and Scoring

The paper outlines specifics for fairness-aware preprocessing and model training methods applicable within the themis-ml framework. Methods such as Relabelling and Reweighting adjust datasets to balance class outcomes, while algorithms like Prejudice Remover Regularizer (PRR) and Additive Counterfactually Fair (ACF) models generate models that minimize discriminatory predictions.

Additionally, scoring mechanisms are proposed for evaluating both group-level and individual-level discrimination in datasets and predictions. Measures such as mean difference (MD), normalized mean difference (NMD), consistency, and situation test score provide quantitative frameworks for assessing the fairness and equity of model predictions.

Evaluation of Fairness-aware Techniques

The evaluation of themis-ml showcases its impact across multiple conditions and model types, underscoring the potential for various fairness-aware techniques to reduce PD predictions. The paper draws attention to the fairness-utility tradeoff, investigating how model fairness is influenced by these techniques' implementation, and to what extent such implementation impacts predictive performance.

Future Directions and Implications

The research acknowledges the need for further exploration into extending fairness-aware interfaces to multi-class and regression settings. Additionally, deeper insights into hyperparameter tuning and the compositional efficacy of combined fairness techniques are warranted.

Moreover, societal challenges are highlighted, particularly that definitions of fairness often exclude marginalized voices, emphasizing the importance of inclusive dialogue in shaping fairness-aware systems. Future developments should seek to establish legal and technical frameworks that foster transparency, accountability, and fair representation in ML applications.

Overall, themis-ml offers a foundational tool for addressing discrimination within ML-driven decision systems, proposing both practical and theoretical expansions that align machine learning's application with equitable social advancement.