Research map / Learning theory and algorithms / Fairness
Identifying Reasons for Bias: An Argumentation-Based Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a model-agnostic, argumentation-based method that builds a weighted argumentation graph over a queried individual's attribute-value pairs and k nearest neighbours, then uses h-Categorizer semantics to find the attribute-value pairs most responsible for a negative classification, as a way to detect individual-level bias.
Problem
Detecting why an individual is classified differently from similar individuals, to identify individual-level unfairness without needing training data, labels, or pre-specified protected attributes.
Approach
Represent attribute-value pairs of a queried individual and its k nearest neighbours as arguments in a quantitative argumentation framework, with attacks from neighbours who have different classifications weighted by the proportion of neighbours sharing each value; the weakest final arguments under weighted h-Categorizer semantics are the reasons for the negative classification.
Key results
- On Adult, 8.2% of queried individuals had (sex, Female) as the weakest argument versus 1.8% for Male, and (race, Black) was the weakest in 245 cases (3.4%).
- On Adapted Adult with injected bias-attr, the method identified bias-attr=0 as the weakest argument for 56.4% of queried individuals (76.8% with 10 neighbours, 85.3% with 15), with no false identification among individuals having differently classified neighbours.
- On Bank Marketing, 8.9% of queried individuals had (age, YoungOrOld) as the weakest argument versus 1.1% for MidAge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the argumentation method to identify weakest arguments among queried individuals | Adult · Adult queried individuals | weakest-argument proportion: sex=Female: 595 (8.2%) | Helped |
| Use the argumentation method to identify weakest arguments among queried individuals | Adult · Adult queried individuals | weakest-argument proportion: race=Black: 245 (3.4%) | Helped |
| Use the argumentation method to identify weakest arguments among queried individuals | Bank Marketing · Bank Marketing queried individuals | weakest-argument proportion: age=YoungOrOld: 112 (8.9%) | Helped |
| Use the argumentation method to identify weakest arguments among queried individuals | Bank Marketing · Bank Marketing queried individuals | weakest-argument proportion across marital values: Married 59 (4.7%); Single 68 (5.4%); Divorced 51 (4.0%) | No clear effect |
| Add synthetic bias attribute and classify by its value | Argumentation-based method · Adapted Adult test set | queried individuals with bias-attr=0 identified among weakest arguments: 56.4% | Helped |
| Increase the number of similar individuals from 5 to 10 (ablation) | Argumentation-based method · Adapted Adult test set | queried individuals identified with bias-attr=0 as reason: 56.4% -> 76.8% | Helped |
| Increase the number of similar individuals from 5 to 15 (ablation) | Argumentation-based method · Adapted Adult test set | queried individuals identified with bias-attr=0 as reason: 85.3% | Helped |
| Use a smaller convergence threshold than ε=0.01 (ablation) | Argumentation-based method · Adult and Bank Marketing | weakest-argument identification | No clear effect |
| Use a larger convergence threshold than ε=0.01 (ablation) | Argumentation-based method · Adult and Bank Marketing | weakest-argument identification | Hurt |
| Use KD clustering instead of Ball-tree clustering (ablation) | Argumentation-based method · Adult and Bank Marketing | performance | No clear effect |
| Use Quadratic Energy Model semantics instead of Hbs (ablation) | Argumentation-based method · Adult and Bank Marketing | weakest-argument identification | No clear effect |
Surprising findings
- Using only individual fairness with consistent neighbours means 43.6% of injected-bias individuals are not flagged, since all neighbours share the same negative label.
Limitations
- Individuals whose similar neighbours all share their classification cannot have bias identified, as the method reduces to individual fairness.
- Results are reported qualitatively as counts of weakest arguments, with no comparison to existing fairness or bias detection methods.
- Only categorical attributes with Hamming distance and two datasets with logistic regression are evaluated.
Experimental setup
Datasets: Adult Bank Marketing
Benchmarks: Adult Bank Marketing
Metrics: count and proportion of weakest arguments by protected attri accuracy F1
Builds on: h-Categorizer semantics weighted argumentation graphs Ball Tree nearest neighbours
Closest papers by meaning
- Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness (NeurIPS 2025 · similarity 0.87)
- Arbitrariness and Social Prediction: The Confounding Role of Variance in Fair Classification (AAAI 2024 · similarity 0.87)
- Size-adaptive Hypothesis Testing for Fairness (NeurIPS 2025 · similarity 0.86)
- Fairness Shields: Safeguarding against Biased Decision Makers (AAAI 2025 · similarity 0.86)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.86)
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.