TL;DR
This paper introduces a novel, model-agnostic approach to identify why an individual is unfairly classified by a machine learning system, focusing on individual fairness rather than group fairness. By representing an individual and their similar counterparts as arguments in a weighted argumentation framework, the method pinpoints specific attribute-value pairs that most contribute to a biased outcome, offering transparent explanations.
What Problem Does This Paper Address?
As machine learning (ML) systems become more integrated into critical decision-making processes (e.g., criminal justice, recruitment), ensuring their fairness is paramount. While much research has focused on group fairness (ensuring fairness across demographic groups) and developing metrics to quantify bias, there's a significant gap in understanding why a specific individual is treated unfairly. Existing methods often require access to sensitive training data, protected attributes, and can be opaque about the root causes of individual bias. This paper tackles the challenge of providing transparent, individual-level explanations for biased classifications, moving beyond simply detecting bias to understanding its origins.
Methods
The core of this paper is a novel model-agnostic, argumentation-based approach for identifying individual bias. Here's how it works:
- Defining Similar Individuals: The method first identifies a set of individuals from the test data that are "similar" to the queried individual. Similarity is determined using a Ball Tree clustering algorithm based on the Hamming distance between their attribute values. Numerical attributes are categorized to fit this framework.
- Constructing a Weighted Argumentation Graph (WAG):
- Arguments: Each unique attribute-value pair present in the queried individual and their similar individuals forms an argument. For instance, if the queried individual has
race=Blackand similar individuals haverace=White, then(race, Black)and(race, White)become arguments.
- Arguments: Each unique attribute-value pair present in the queried individual and their similar individuals forms an argument. For instance, if the queried individual has
- Attacks: An attack is established from an attribute-value pair of a similar individual to an attribute-value pair of the queried individual if:
- The similar individual has a different classification than the queried individual.
- The attribute-value pair of the similar individual differs from that of the queried individual for a specific attribute.
- Attack Strengths: The strength of an attack is proportional to the number of similar individuals who share a particular attribute-value pair that differs from the queried individual's. This reflects how prevalent a specific difference is among similar individuals.
- Initial Weights: All arguments are initially assigned a weight of 1.
- Calculating Final Argument Weights: The paper employs the Weighted h-Categorizer (Hbs) semantics to calculate the final weights of arguments. This semantics is chosen because it can handle cycles, accounts for both the strength and quantity of attacks, and is efficient. The Hbs semantics iteratively updates argument weights, reducing them based on incoming attacks. The final weight of an argument represents its "strength" or "acceptability" after considering all conflicting arguments.
- Identifying Bias Reasons: The attribute-value pairs with the smallest final weights are identified as the weakest arguments. These weakest arguments are considered the primary reasons contributing to the queried individual's negative classification, thus explaining the bias.
Datasets
The paper evaluates its method on two commonly used datasets:
- Adult Dataset: A widely used dataset for fairness research, containing demographic and socio-economic information.
- Bank Marketing Dataset: Another standard dataset for evaluating classification models, often used in fairness contexts.
The authors also introduce synthetic bias into the Adult dataset to specifically test their method's ability to detect known biases.
Key Insights and Ideas
- Individual Fairness Focus: The paper shifts the focus from group fairness metrics to understanding the specific reasons behind an individual's unfair treatment. This is crucial because individuals are primarily concerned with their personal outcomes.
- Argumentation as Explanation: Computational argumentation provides a natural and transparent way to represent conflicting evidence (attribute differences) and derive a reasoned conclusion about the cause of a classification.
- Model-Agnosticism: The approach does not depend on the underlying ML model, making it broadly applicable to any binary classifier.
- No Protected Attributes Required (for the method itself): While the evaluation uses protected attributes to demonstrate bias, the core method of constructing the argumentation graph and calculating weights does not require prior specification of these attributes. This is a significant advantage for real-world scenarios where such data might be unavailable or sensitive.
- Weakest Argument = Reason for Bias: The core insight is that the attribute-value pairs with the lowest final weights in the argumentation framework are the most influential factors driving a biased classification.
Key Results
- Adult Dataset: The method successfully identified attribute-value pairs that correlate with known biases. For instance, it highlighted
sex=Femaleas a contributing factor to negative classifications in 8.2% of queried individuals, aligning with the dataset's known gender disparities. Similarly,race=Blackwas identified as a stronger contributor to negative classifications thanrace=White. - Bank Marketing Dataset: The approach identified
age=YoungOrOldas a significant factor in negative classifications for 8.9% of queried individuals, consistent with observed patterns. - Synthetic Bias: When synthetic bias (
bias-attr=0) was introduced into the Adult dataset, the method correctly identifiedbias-attr=0as the weakest argument in a significant proportion of cases (56.4% when at least one similar individual was positively classified). - Scalability: The method was found to be scalable, with experiments on large datasets completing in under 2 minutes.
Limitations and Future Work
- Limitation of Individual Fairness: The paper acknowledges a limitation where if all similar individuals have the same classification (even if negative), no bias is detected for the queried individual, as they are treated "consistently." This highlights that individual fairness alone might not capture all forms of bias.
- Future Work:
- Extend to More Datasets and Semantics: Explore other datasets and argumentation semantics to ensure robustness and identify more subtle causes of bias.
- Natural Language Explanations: Develop templates to translate the identified attribute-value pairs into more intuitive, human-readable explanations.
- Group Fairness Integration: Investigate how to better integrate group fairness considerations or explore scenarios where both individual and group fairness are addressed.
- Alternative Similarity Measures: Experiment with different definitions of individual similarity.
Summary updated