Fair Classification with Partial Feedback: An Exploration-Based Data Collection Approach

Overview

• Predictive models, like those used for credit lending decisions, often have limited data on negative outcomes due to past biases in classification.
• This can lead to erroneous models that perpetuate those biases.
• The paper presents an approach to train a classifier while actively collecting data on underrepresented subpopulations to improve the model.
• The approach provides guarantees around exploring all subpopulations, bounding false positives, and converging to a "desired" classifier.
• The exploration strategy can be chosen to improve learning and incorporate fairness considerations.
• Evaluations show this approach boosts data quality and true positive rates across groups, with only a small reduction in predictive utility.

Plain English Explanation

In many predictive tasks, such as deciding who should be granted a loan, the data available for training models is often biased. This is because the model only has information about people who were approved for loans in the past, while knowing little about those who were denied.

This lack of information on "negative" outcomes can lead the model to make inaccurate predictions, potentially perpetuating the same biases present in the historical data.

The researchers in this paper propose an approach to address this issue. Their method trains a classifier using the available data, but also includes strategies to actively collect outcome data on subgroups that would otherwise be overlooked.

This "exploration" process comes with important guarantees - it ensures all subpopulations are examined, limits the fraction of false positive classifications, and ensures the final trained model converges to a "desired" classifier.

Importantly, the exploration strategy can be tailored to the specific context, allowing incorporation of fairness considerations. For example, the exploration could prioritize underrepresented groups to improve equity.

Evaluations on real-world datasets show this approach consistently improves the quality of the collected data and increases the true positive rate across different groups, with only a small reduction in the model's overall predictive performance.

Technical Explanation

The key technical insight of this paper is that in many predictive contexts, the training data available to machine learning models is inherently biased due to past decisions to grant or deny access to the "true" outcome information.

For example, in credit lending, the model only has data on the repayment behavior of applicants who were approved for loans in the past. It lacks information on the outcomes of those who were incorrectly denied loans previously.

This "selective labels" problem can lead to erroneous classifiers that perpetuate historical biases. To address this, the authors propose an approach that trains a classifier using the available data, but also includes a family of exploration strategies to actively collect outcome data on underrepresented subpopulations.

The exploration strategies come with several key guarantees:

1. All subpopulations will be explored to some degree, ensuring comprehensive data collection.
2. The fraction of false positive classifications will be bounded, limiting harm.
3. The trained classifier will converge to a "desired" classifier, defined by the specific context and fairness objectives.

The right exploration strategy is context-dependent and can be chosen to improve learning guarantees and incorporate group fairness properties. For example, the exploration could prioritize data collection on disadvantaged groups to improve equitable model performance.

Evaluations on real-world datasets show this approach consistently improves the quality of the collected data and the true positive rate across different groups, with only a small reduction in overall predictive utility. This demonstrates the value of the proposed method in addressing the selective labels problem and training fairer, more accurate predictive models.

Critical Analysis

The paper presents a compelling technical approach to address the selective labels problem, which is a significant challenge in many real-world predictive contexts. The theoretical guarantees around exploration, false positive rates, and model convergence are particularly valuable.

However, the authors acknowledge several limitations and areas for further research. For example, the approach assumes the existence of a "desired" classifier, which may be difficult to define in practice, especially when considering complex notions of fairness. Differentially Private Fair Binary Classifications and Intrinsic Fairness-Accuracy Tradeoffs under Equalized Odds explore related challenges in defining and achieving fairness in predictive models.

Additionally, the evaluation is limited to a few real-world datasets, and the authors note that the effectiveness of the exploration strategies may depend on the specific context and dataset characteristics. Further research is needed to understand how the approach generalizes to a wider range of applications and data environments.

Another potential concern is the practical implementation and deployment of the proposed exploration strategies. While the theoretical foundations are solid, the operational details of how to collect additional outcome data in a real-world setting may present significant challenges, such as obtaining user consent, managing ethical and legal considerations, and ensuring the exploration process does not cause unintended harm.

Overall, this paper presents an important and technically rigorous contribution to the growing body of research on addressing selective labels and fairness in predictive modeling. The proposed approach offers a promising direction, but further work is needed to fully understand its broader applicability and practical implications.

Conclusion

This paper tackles the critical challenge of selective labels in predictive modeling, where training data is inherently biased due to past decisions to grant or deny access to "true" outcome information. The authors present an approach that trains a classifier using available data, while also including exploration strategies to actively collect outcome data on underrepresented subpopulations.

The key strengths of this approach are the theoretical guarantees around exploring all subpopulations, bounding false positive rates, and converging to a "desired" classifier. Importantly, the exploration strategy can be tailored to the specific context, allowing for the incorporation of fairness considerations.

Evaluations on real-world datasets demonstrate the method's ability to consistently improve the quality of collected data and increase true positive rates across different groups, with only a small reduction in overall predictive utility. This highlights the value of the proposed technique in addressing selective labels and training fairer, more accurate predictive models.

While the paper presents a compelling technical contribution, further research is needed to fully understand the broader applicability of the approach, the practical challenges of implementation, and the nuances of defining "desired" classifiers in complex real-world settings. Nonetheless, this work represents an important step forward in the ongoing effort to develop more equitable and reliable predictive systems.