Fairness through Instance Level Data Attribution and Pruning

This direction is split into other directions: influence-based data pruning (influence-based bias mitigation). Its last version stays readable and citable here.

Version 1 of 158 papers99 conceptsfrozen 2026-10-04

In the traditional pursuit of algorithmic fairness, researchers have largely focused on "in-processing" techniques: modifying model architectures or adding fairness constraints to loss functions. However, these methods often face a stubborn "fairness-accuracy tax," where improving equity inevitably degrades predictive performance.

A growing body of research, led by researchers like Yang Liu, Hongfu Liu, and Jialu Wang, suggests a different path. Instead of fixing the model, they propose fixing the data. By identifying and removing the specific training instances that drive discriminatory behavior, practitioners can build models that are both fairer and, surprisingly, often more accurate.

The Intuition: Bias is Not Uniform

The core insight of this research direction is that model bias is rarely a property of the entire dataset. Instead, it is often concentrated in a small subset of "harmful" or "influential" training points. These points might contain biased human labels, reinforce spurious correlations, or represent extreme stereotypical associations.

As Anshuman Chhabra and Hongfu Liu argue in "What Data Benefits My Classifier?", if we can pinpoint these specific samples, we can curate a training set that guides the model toward more equitable decision boundaries without the need for complex architectural changes.

Measuring Harm with Influence Functions

To curate data effectively, we need a way to measure how much a single training point $(x_i, y_i)$ affects a model’s fairness. The primary tool for this is the influence function.

Originally a tool from robust statistics, influence functions allow researchers to estimate how a model’s parameters—and consequently its fairness metrics—would change if a specific data point were removed, all without actually retraining the model. Jialu Wang and Yang Liu demonstrated that by calculating the gradient of a fairness objective (like demographic parity) with respect to training samples, they could identify points that disproportionately drive bias.

This approach is highly efficient. For example, Yushun Dong and colleagues developed the BIND framework for Graph Neural Networks, which uses influence-based attribution to identify harmful nodes with a 450x speedup over traditional retraining methods.

Surgical Pruning and Reweighing

Once the "harmful" points are identified, the strategy is simple: remove them or downweight them.

  • Data Pruning: Removing a small fraction of the most biased points—sometimes as little as 10%—can achieve fairness gains comparable to specialized debiasing algorithms. In vision-language models, Eric Slyman and colleagues introduced FairDeDup, showing that semantic deduplication can be made fairness-aware to preserve minority representation while cutting dataset size in half.
  • Data Reweighing: Instead of deletion, Peizhao Li and Hongfu Liu proposed a reweighing scheme that granularly adjusts the importance of each sample. Their method aims for "cost-free fairness," breaking the traditional trade-off by identifying points that harm fairness without contributing to predictive utility.

Why It Works: The Win-Win Scenario

The most compelling finding in this field is that data-centric fairness often leads to a "win-win" for accuracy and equity. This happens because the data points that induce bias are frequently the same points that are noisy, mislabeled, or rely on non-generalizable features.

For instance, Yunyi Li and Maytal Saar-Tsechansky showed that by removing instances where human labels contradict learned demographic patterns, they could eliminate "label bias." Similarly, Yixuan Zhang found that filtering for both utility and fairness allows models to converge faster and reach higher accuracy than training on the full, unfiltered dataset.

Selective Classification as Distributional Pruning

A related mechanism for achieving fairness is selective classification or abstention. Instead of pruning the training set, frameworks like FAN (Fair Abstention classifier with No harm), developed by Tongxin Yin and Yang Liu, allow a model to "abstain" from making a prediction on samples where it cannot guarantee fairness or accuracy.

Mathematically, this acts as a form of post-hoc pruning. By strategically withholding predictions on high-disparity samples, the model effectively reshapes its output distribution to satisfy parity constraints on the remaining subset, ensuring that no sub-population is worse off than they were with the original model.

Summary

Data-centric fairness shifts the focus from the "how" of learning to the "what" of learning. By treating fairness as a data quality problem, researchers have shown that we can audit datasets at the instance level, attribute bias to specific points, and surgically refine our training sets to build AI systems that are both more equitable and more robust.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_fairness_through_instance_level_data_att_66bef350,
  title = {Fairness through Instance Level Data Attribution and Pruning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fairness-through-instance-level-data-attribution-and-pruning/rdr_66bef350ddeb49c1b22cee8caacb8071},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_fairness_through_instance_level_data_att_66bef350_v1,
  title = {Fairness through Instance Level Data Attribution and Pruning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fairness-through-instance-level-data-attribution-and-pruning/rdr_66bef350ddeb49c1b22cee8caacb8071/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}