When machine learning models exhibit bias—whether predicting higher loan default rates for minority groups or associating specific genders with certain professions—the traditional remedy has been model-centric. Practitioners typically add fairness constraints to the loss function, modify the model architecture, or adjust the final classification thresholds. However, these methods treat the underlying dataset as an immutable block, often forcing a steep trade-off where overall predictive accuracy drops significantly.
A growing body of research shifts this paradigm from model-centric to data-centric. Instead of forcing a model to learn around biased data, why not surgically clean the dataset itself? As argued by Anshuman Chhabra et al. in "What Data Benefits My Classifier?" Enhancing Model Performance and Interpretability through Influence-Based Data Selection, we can keep the model fixed and curate the data instead. The core insight driving this direction is that algorithmic unfairness is rarely distributed evenly across a dataset. Instead, bias is often concentrated within a small subset of highly influential training points. By identifying and pruning or downweighting these specific instances, we can achieve dramatic fairness gains with negligible impact on overall utility.
The Core Insight: Concentrated Bias
Historically, auditing datasets for systemic bias has been an elusive challenge. As Jialu Wang et al. note in Understanding Instance-Level Impact of Fairness Constraints, and Saachi Jain et al. echo in Improving Subgroup Robustness via Data Selection, there has been a significant lack of tools to perform instance-level data auditing. This has left practitioners blind to exactly which training examples drive a model toward discriminatory predictions or spurious shortcuts.
Data-centric fairness frameworks solve this by treating unfairness as a data quality issue rather than an inherent mathematical conflict. Frameworks like AIM, developed by Zhining Liu et al. in AIM: Attributing, Interpreting, Mitigating Data Unfairness, establish a transparent audit trail by identifying specific biased samples through comparisons with similar peers across different demographic groups. Strikingly, Yang Liu et al. demonstrate in Understanding Unfairness via Training Concept Influence that modifying or removing less than ten percent of these highly influential samples can yield fairness improvements competitive with complex in-processing algorithms.
The Mechanism: Calculating Fairness Influence
To surgically remove the data points driving unfairness, we must first quantify each instance's impact on the model's fairness metrics. Retraining a model from scratch to test the removal of every single data point is computationally impossible for deep networks. Instead, researchers leverage influence functions from robust statistics to approximate this impact.
Intuitively, an influence function calculates how a tiny change in the weight of a training instance $z$ shifts the model's optimal parameters $\theta$. By using the first derivative (gradients) of the instance's loss and the second derivative (the Hessian matrix) of the total training loss, we can map out the local geometry of the loss surface.
To optimize for equity, the authors of Understanding Instance-Level Impact of Fairness Constraints adapted these influence functions to account directly for fairness constraints (such as demographic parity or equal opportunity). Training instances that exhibit low or negative influence on these fairness constraints are the primary sources of algorithmic bias. Removing them allows the model's decision boundary to naturally shift toward an equitable position.
This mathematical framework has been extended to handle increasingly complex scenarios:
- Virtual Removal without Refitting: The Fair Infinitesimal Jackknife (Fair-IJ) framework, introduced by Prasanna Sattigeri et al. in Fair Infinitesimal Jackknife: Mitigating the Influence of Biased Training Data Points Without Refitting, utilizes scalable Hessian approximations to update model weights as if biased data points were removed, bypassing the need to refit the model entirely.
- Cross-Loss Objectives: Standard influence functions require the training and testing objectives to be identical. To debias models where this isn't the case—such as self-supervised word embeddings—Andrew Silva et al. developed Cross-Loss Influence Functions (CLIF) in Cross-Loss Influence Functions to Explain Deep Network Representations, allowing practitioners to trace the impact of training samples on arbitrary downstream test metrics.
- Structured Data: This instance-level attribution is highly versatile, extending to relational structures like Graph Neural Networks via the BIND framework proposed by Song Wang in Interpreting Unfairness in Graph Neural Networks via Training Node Attribution, and to knowledge graphs, where Yupei Du et al. trace stereotypical associations back to specific relational triples in Understanding Gender Bias in Knowledge Base Embeddings.
Breaking the Fairness-Utility Trade-off
Why does influence-based data pruning circumvent the traditional fairness-accuracy trade-off?
When we apply global optimization constraints to a model during training, we restrict its entire hypothesis space, which often degrades performance on clean, legitimate data patterns. In contrast, data-centric pruning isolates and removes instances that represent historical bias, label noise, or uninformative demographic correlations. As Peizhao Li et al. demonstrate in Achieving Fairness at No Utility Cost via Data Reweighing with Influence, downweighting these problematic points can achieve "cost-free fairness"—improving equity while maintaining or even boosting predictive utility.
This targeted approach yields massive efficiency gains. The authors of Understanding Instance-Level Impact of Fairness Constraints show that pruning data based on fairness influence scores can achieve an optimal balance of accuracy and equity using as little as 20% to 40% of the original dataset. Similarly, in decentralized settings, the Frog framework by Yejia Liu et al. in Enabling SQL-based Training Data Debugging for Federated Learning allows users to issue SQL-based complaints about model behavior, identifying and removing influential training records securely across federated nodes with negligible performance loss.
Active Sampling and Alignment
The principles of influence-based data selection apply not only to pruning existing datasets but also to acquiring new data and aligning large language models.
Instead of filtering a fully labeled dataset, Fair Influential Sampling (FIS), designed by Jinlong Pang et al. in Fair Classifiers Without Fair Training: An Influence-Guided Data Sampling Approach, uses gradient alignment to select the most equitable data points from an unlabeled pool. By matching unlabeled examples with the gradients of a small, fair validation set, the framework systematically mitigates risk disparity during data acquisition.
This philosophy of prioritizing data quality over raw scale is echoed across multiple domains:
- Active Sampling in NLP: Lele Sha et al. utilize active sampling in Bigger Data or Fairer Data? Augmenting BERT via Active Sampling for Educational Text Classification to select demographically uninformative data for language models, achieving a 52% improvement in fairness.
- Parametrized Resampling: Vladimiro González-Zelaya et al. introduce continuous parameters in Optimising Fairness Through Parametrised Data Sampling to resample data to meet target positive rates, optimizing fairness without altering the underlying classification algorithm.
- Label Bias Mitigation: YIXUAN ZHANG et al. leverage zero-shot predictors and peer-prediction in Navigating Towards Fairness with Data Selection to filter out samples affected by label bias, accelerating model convergence while reducing fairness violations.
- Simple Modifications: Erenay Dayanik et al. show in Masking Actor Information Leads to Fairer Political Claims Detection that simple data-level modifications, like masking sensitive attributes or actor names, can outperform complex adversarial training in balancing fairness and performance.
- Computer Vision Scale: Douwe Kiela demonstrates in Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision that uncurated, high-volume datasets often embed deeper societal stereotypes than smaller, well-managed datasets, proving that curation outperforms raw scale.
Lately, this data-centric lens has become vital for aligning generative models. While Reinforcement Learning from Human Feedback (RLHF) is widely used to make models helpful, Aaron Jiaxun Li demonstrates in More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness that standard preference alignment can inadvertently increase stereotypical bias and decrease truthfulness. By utilizing data attribution scores, practitioners can identify and prune the specific preference examples driving these negative behaviors. Furthermore, Sheng Ouyang et al. show in Towards Reward Fairness in RLHF: From a Resource Allocation Perspective that optimizing reward models for fairness prevents them from inflating scores for biased features (like response length), ultimately allowing downstream policies to achieve superior performance using smaller, cleaner training sets.
Key Foundational Work
Certifying and removing disparate impact (2014) — contributed 1 concept elements to this direction.
Masking Actor Information Leads to Fairer Political Claims Detection (2020) — contributed 1 concept elements to this direction.