Vijay Keswani,L. Elisa Celis
Abstract:In real-world classification settings, such as loan application evaluation or content moderation on online platforms, individuals respond to classifier predictions by strategically updating their features to increase their likelihood of receiving a particular (positive) decision (at a certain cost). Yet, when different demographic groups have different feature distributions or pay different update costs, prior work has shown that individuals from minority groups often pay a higher cost to update their features. Fair classification aims to address such classifier performance disparities by constraining the classifiers to satisfy statistical fairness properties. However, we show that standard fairness constraints do not guarantee that the constrained classifier reduces the disparity in strategic manipulation cost. To address such biases in strategic settings and provide equal opportunities for strategic manipulation, we propose a constrained optimization framework that constructs classifiers that lower the strategic manipulation cost for minority groups. We develop our framework by studying theoretical connections between group-specific strategic cost disparity and standard selection rate fairness metrics (e.g., statistical rate and true positive rate). Empirically, we show the efficacy of this approach over multiple real-world datasets.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to reduce the cost differences among different groups during strategic feature manipulation in fair classification. Specifically, different social groups pay different costs when they try to change their own features to obtain positive decisions (such as loan approval, content release, etc.) due to historical biases or cost differences. The paper points out that traditional fairness constraints (such as statistical rate fairness, equal opportunity, etc.) can reduce the inequality in the prediction performance of classifiers, but cannot guarantee the reduction of the differences in strategic manipulation costs among these groups. Therefore, the author proposes a new optimization framework, aiming to construct a classifier that can both reduce the selection rate difference and the strategic manipulation cost difference, so as to provide more equal opportunities for all groups.
### Core Problems of the Paper
1. **Differences in Strategic Manipulation Costs**: There are differences in the costs paid by different groups when they try to change their own features to influence the classifier's decision, especially minority groups often need to pay higher costs.
2. **Limitations of Traditional Fairness Constraints**: Existing fairness constraints (such as statistical rate fairness, equal opportunity, etc.) can reduce the prediction inequality of classifiers to a certain extent, but cannot effectively reduce the differences in strategic manipulation costs.
3. **Constructing a Fair Classifier**: How to design a classifier so that it can reduce the differences in strategic manipulation costs among different groups while satisfying traditional fairness constraints.
### Solutions
The paper proposes a new optimization framework. By introducing constraints on the differences in strategic manipulation costs, a more fair classifier is constructed. The specific steps include:
1. **Theoretical Analysis**: Research on the relationship between standard selection rate fairness indicators (such as statistical rate, true positive rate difference) and the differences in strategic manipulation costs.
2. **Constructing a Classifier**: Use the results of theoretical analysis to design a classifier that can simultaneously reduce the selection rate difference and the strategic manipulation cost difference.
3. **Empirical Analysis**: Verify the effectiveness of the proposed method through experiments on multiple real - data sets.
### Main Contributions
1. **Theoretical Analysis**: It is proved that even when using the selection rate fairness constraint, the classifier may still produce large differences in strategic manipulation costs.
2. **Optimization Framework**: Propose an optimization framework that combines selection rate fairness and differences in strategic manipulation costs.
3. **Empirical Results**: Demonstrate the effectiveness of the proposed method on the FICO credit data set and the Adult income data set.
### Related Work
The paper also reviews research in related fields, including early research on differences in strategic manipulation costs, methods for evaluating classifier fairness in strategic settings, and how to consider the action ability of individuals in classifiers. These studies provide theoretical basis and methodological support for this paper.
### Conclusion
Through theoretical analysis and empirical research, the paper shows how to reduce the differences in strategic manipulation costs among different groups in fair classification, thereby providing more equal opportunities for all groups. This not only helps to improve the fairness of classifiers, but also promotes social fairness and justice.