Conscientious Classification: A Data Scientist's Guide to Discrimination-Aware Classification (1907.09013v1)

Abstract: Recent research has helped to cultivate growing awareness that machine learning systems fueled by big data can create or exacerbate troubling disparities in society. Much of this research comes from outside of the practicing data science community, leaving its members with little concrete guidance to proactively address these concerns. This article introduces issues of discrimination to the data science community on its own terms. In it, we tour the familiar data mining process while providing a taxonomy of common practices that have the potential to produce unintended discrimination. We also survey how discrimination is commonly measured, and suggest how familiar development processes can be augmented to mitigate systems' discriminatory potential. We advocate that data scientists should be intentional about modeling and reducing discriminatory outcomes. Without doing so, their efforts will result in perpetuating any systemic discrimination that may exist, but under a misleading veil of data-driven objectivity.

Conscientious Classification: Addressing Discrimination in ML Models

The paper "Conscientious Classification: A Data Scientist's Guide to Discrimination-Aware Classification" by d'Alessandro, O'Neil, and LaGatta critically examines the subtle implications of ML systems in perpetuating societal biases. It emphasizes the need for the data science community to account for discrimination within their modeling processes, asserting that advances in ML can inadvertently reinforce historical biases under the guise of objectivity.

The authors discuss the distinction between disparate treatment and disparate impact, exploring legal definitions of discrimination. Disparate treatment involves differential treatment based on membership in a protected class, while disparate impact entails differential outcomes without explicit consideration of class membership, often through correlated factors. These definitions become pivotal when assessing ML models' fairness and societal impact.

The paper dissects the elements within the typical data science model development lifecycle using the CRISP-DM process framework, identifying potential points where discrimination can emerge. These include:

1. Data Issues: Discriminatory biases can be embedded within the datasets used for training models. The society from which data is derived may have systemic biases that are mirrored in the models unless corrective measures are taken.
2. Misspecification: This can occur through either misuse of feature sets or inappropriate modeling approaches—such as proxy targeting or improper cost function settings—which may exacerbate discriminatory outcomes.
3. Process Failures: Lack of proper auditing and feedback mechanisms can lead to perpetuated biases, particularly if interventions influenced by the model outcomes alter the data distribution in a manner that exacerbates existing biases.

The authors advocate for a proactive adoption of discrimination-aware data mining techniques across pre-processing, in-processing, and post-processing stages to mitigate these biases. They suggest implementing discrimination metrics to develop unit tests for potential biases in the data and model predictions, complemented by frameworks that marry legal and ethical considerations with model development.

Furthermore, the paper explores case studies, notably predictive policing and applicant screening systems, to illustrate the complex interplay between ML systems and societal discrimination. These case studies highlight the risks of proxy target variables and feedback loops that reinforce biases, underlining the necessity for rigorous auditing and intervention techniques.

In practice, the paper suggests that data scientists should remain highly sensitive to misclassification costs, considering both statistical significance and practical impact. The inclusion of fairness regularizers in cost functions and leveraging open-source libraries for pre/post-processing measures are recommended strategies to enhance accuracy while reducing discriminatory impacts.

The implications of this work are pivotal for both theoretical advancements and practical applications in AI, as it calls for data scientists to recognize their ethical and legal responsibilities in model development. The propagation of discriminatory biases through ML systems poses significant risks, which can be mitigated through conscientious efforts to integrate fairness into every stage of the modeling process. Future developments in AI should continue to emphasize discrimination-aware practices, ensuring models contribute positively to societal equity rather than amplifying entrenched disparities.