Team research project · 2026
Fairness interventions and model explanations
Looking beyond prediction scores to understand how fairness interventions change decisions and feature attributions.
Conceptual illustration; not an experimental result.
Individual contribution
My role
My work covered fairness metric implementation, reweighing, visualisation and explainability analysis. These contributions formed part of a shared study of model behaviour and decision-level interventions.
Contribution statement in the reportResearch question
A fairer set of predictions does not necessarily mean a model has learned a different representation. This project studies that distinction on the Adult Census Income dataset, connecting group fairness measurements with local and global explanations.
Methodology
The team compared logistic regression and random forest models, using reweighing and threshold adjustment as two different intervention points. Demographic parity and equal opportunity assess outcomes; SHAP and LIME help inspect the patterns behind them. An Explanation Fairness Gap compares feature attributions between gender groups.
Findings
The reports distinguish changes in model behaviour from changes in the decision rule. Reweighing affects the fitted model, while threshold adjustment acts on its output. Reading explanations alongside fairness measures makes that difference visible.
Limitations
This is a course research project on one benchmark dataset, not evidence that a system is universally fair. Fairness depends on the metric, group definition and task. The two reports below contain the experimental setup, findings and limitations.
Related research
Conditioning diffusion transformers