Writing study: Revisiting (Un)Fairness in Recourse by Minimizing Worst-Case Social Burden
← AAAI 2026 oral writing corpus
Ainhize Barrainkua, Giovanni De Toni, Jose A. Lozano, Novi Quadrianto · Official AAAI 2026 oral record
Figure, table and section ledger
Editorial contribution type: Training, objectives & optimization (primary).
Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.
- Canonical PDF
- 9 physical pages
- Before references / appendix
- ≈ 7.22 page-equivalents
- Main figures / tables
- 2 / 2
- Attached appendix figures / tables
- 0 / 0
- Attached appendix span
- No boundary detected; not assumed zero
Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.
Source-bound correction: Checked 2026-10-25. This source numbers its main sections; subsection numbers and the two formal definitions are not extra top-level sections. Figures 1–2 and Tables 1–2 are counted. Algorithm 1 is retained in the writing study, not counted as a figure or table by this two-kind caption metric. Physical pp.2,6,7 were visually inspected. External appendices and code are outside the measurement scope.
Section-by-section lengths and source pages
The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.
| Top-level heading | Starts on PDF page | Estimated span | Share of entire PDF |
|---|---|---|---|
| 1 Introduction | 1 | ≈ 1.54 pp. | 17.11% |
| 2 Preliminaries and Related Work | 2 | ≈ 1.08 pp. | 12% |
| 3 Revisiting (Un)Fairness in Recourse | 3 | ≈ 1.43 pp. | 15.89% |
| 4 MISOB: MInimax SOcial Burden | 5 | ≈ 0.91 pp. | 10.11% |
| 5 Empirical Evaluation | 6 | ≈ 1.13 pp. | 12.56% |
| 6 Discussion and Limitations | 7 | ≈ 0.41 pp. | 4.56% |
| 7 Conclusions | 7 | ≈ 0.24 pp. | 2.67% |
| Acknowledgments | 8 | ≈ 0.18 pp. | 2% |
| References | 8 | ≈ 1.78 pp. | 19.78% |
All detected figure and table locations
Attached appendix
No recognized caption labels.
Argument and organization
- 1 Introduction; 2 Preliminaries and Related Work
Motivates the difference between cost parity and who needs recourse, then defines the classifier, recourse operation and related fairness criteria.
- 3 Revisiting (Un)Fairness in Recourse
Separates three defects in the conventional framing: classifier decisions, false negatives and gap-only comparisons. Definitions make the changed population conditioning explicit.
- 4 MISOB: MInimax SOcial Burden; 5 Empirical Evaluation
Turns burden estimates into iterative training weights, then tests task performance, group outcomes and sensitivity to the weighting parameter.
- 6 Discussion and Limitations; 7 Conclusions; Acknowledgments; References
Separates unproved convergence and realistic-cost extensions from completed observations, before final production back matter.
Problem, claim and contribution
Uses a two-stage decision-and-recourse example to show why equal per-recipient costs can coexist with unequal population burdens. The contribution list points separately to the formal characterization, critique of cost parity, training procedure and evaluation.
Introduction and Figure 1, physical pp.1–2; contribution list, p.2 · Main paper
Argument and section order
Organizes Section 3 around three specific shortcomings instead of introducing the algorithm immediately. Definitions 3.1–3.2 distinguish all negative predictions from qualified individuals receiving false negatives before Section 4 introduces weighted retraining.
Sections 3.1–3.3 and Definitions 3.1–3.2, pp.4–5; Algorithm 1 and Equation 4, p.5 · Main paper
Figures, captions and visual explanation
Figure 1 pairs a two-stage conceptual diagram with a numerical contrast between conventional and population-aware costs. Figure 2 uses aligned method-by-sensitive-attribute panels and separates bounded task/rate metrics from log-scale cost and burden panels; its caption identifies the weighting sweep, ten-run aggregation and accuracy trade-off.
Figure 1, p.2; Figure 2, p.7 · Main paper
Tables, comparisons and evidence
Table 1 distinguishes what a fairness metric measures from what an improvement method must know. Table 2 groups recourse methods within race, gender and intersectional evaluation, labels metric directions and retains cases where smaller group gaps do not coincide with lower worst-group burden. Its caption reports means and standard deviations over ten runs with alpha fixed to 0.3; the setup identifies random train/test splits.
Table 1, p.3; Table 2 and Experimental Setting, p.6; Results, pp.6–7 · Main paper
Appendix and reproducibility
Routes proofs, experimental configuration and additional models/datasets to separate appendices while keeping the definitions, training pseudocode and central ADULT comparison in the proceedings body. Only the routing statements were inspected, not the linked extended manuscript or supporting claims.
Opening of Section 3, pp.3–4; Section 5 opening and Experimental Setting, p.6; extended-version link, p.1 · Main paper
Clarity and writing conventions
The discussion explicitly leaves convergence analysis and more realistic cost models to future work. However, Equation 7 prints a minimum for worst cost and burden, while Section 3.3 defines worst-group burden using a maximum. This inconsistency should be flagged rather than silently copied into a new metric description.
Section 3.3, p.5; Equation 7, p.6; Discussion and Limitations, p.7 · Main paper
What to reuse—and what not to infer
Transferable lesson: Explain a measurement failure with a concrete two-stage example before introducing the replacement objective, then report absolute group outcomes beside parity gaps so the reader can inspect the distinction.
Boundary: The main experiments use ADULT with a feed-forward classifier; additional-dataset statements point outside the inspected PDF. The printed min/max inconsistency in Equation 7 remains unresolved. No sensitive-attribute access in training or inference does not mean attributes are unnecessary for the reported fairness evaluation. The weighting procedure is empirically motivated here, not a verified convergence or universal fairness guarantee.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-25
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-oriented scan of physical pp.1–7, covering the framing, section progression, definitions, training explanation, captions, evidence and discussion; acknowledgments/reference boundary on p.8 and reference continuation through p.9 checked. Full-page rasters of pp.2, 6 and 7 were inspected. Not exhaustive proof verification, legal analysis, source-code inspection or experimental reproduction.
- Appendix / supplement
- The proceedings PDF contains no attached appendix. Main-text pointers to Appendix A for proofs, B for setup and C for further experiments were inspected. The separately linked extended version and code were not retrieved or inspected for this study; their contents are outside the claimed coverage.
- Visually inspected PDF pages
- 2, 6, 7
This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.