Preferential Sampling

Description

A preprocessing fairness technique developed by Kamiran and Calders that addresses dataset imbalances by re-sampling training data with preference for underrepresented groups to achieve discrimination-free classification. This method modifies the training distribution by prioritising borderline objects (instances near decision boundaries) from underrepresented groups for duplication whilst potentially removing instances from overrepresented groups. Unlike relabelling approaches, preferential sampling maintains original class labels whilst creating a more balanced dataset that prevents models from learning biased patterns due to skewed group representation.

Example Use Cases

Fairness

Preprocessing hiring datasets by preferentially sampling candidates from underrepresented gender and ethnic groups, particularly focusing on borderline cases near decision boundaries, to ensure fair representation whilst maintaining original qualifications and labels for unbiased recruitment model training.

Reliability

Balancing medical training datasets by oversampling patients from underrepresented demographic groups to ensure reliable diagnostic performance across all populations, preventing models from exhibiting reduced accuracy for minority patient groups due to insufficient training examples.

Transparency

Creating transparent credit scoring datasets by documenting and adjusting the sampling process to ensure equal representation across demographic groups, providing clear evidence to regulators that training data imbalances have been addressed without altering original creditworthiness labels.

Limitations

Oversampling minority groups can cause overfitting to duplicated examples, particularly when borderline instances are repeatedly sampled, potentially reducing model generalisation.
Undersampling majority groups may remove important examples that contain valuable information, potentially degrading overall model performance.
Does not address inherent algorithmic bias in the learning process itself, only correcting for representation imbalances in the training data.
Selection of borderline objects requires careful threshold tuning and may be sensitive to the choice of distance metrics or similarity measures used.
May not address intersectional fairness issues when multiple protected attributes create complex group combinations that require nuanced sampling strategies.

Resources

Data preprocessing techniques for classification without discrimination

Research Paper•Faisal Kamiran and Toon Calders•Jun 1, 2012

Classification with no discrimination by preferential sampling

Research Paper•Faisal Kamiran and Toon Calders•May 27, 2010

A Survey on Bias and Fairness in Machine Learning

Documentation•Ninareh Mehrabi et al.•Aug 25, 2019

Related Techniques

Name	Description	Assurance Goals
Confidence Thresholding	Confidence thresholding creates decision boundaries based on model uncertainty scores, routing predictions into different handling workflows depending on their confidence levels. High-confidence predictions (e.g., above 95%) proceed automatically, whilst medium-confidence cases (e.g., 70-95%) may trigger additional validation or human review, and low-confidence predictions (below 70%) receive extensive oversight or default to conservative fallback actions. This technique enables organisations to maintain automated efficiency for clear-cut cases whilst ensuring appropriate human intervention for uncertain decisions, balancing operational speed with risk management across safety-critical applications.	Safety Reliability Transparency
SHapley Additive exPlanations	SHAP explains model predictions by quantifying how much each input feature contributes to the outcome. It assigns an importance score to every feature, indicating whether it pushes the prediction towards or away from the average. The method systematically evaluates how predictions change as features are included or excluded, drawing on game theory concepts to ensure a fair distribution of contributions.	Explainability Fairness Reliability
Saliency Maps	Saliency maps are visual explanations for image classification models that highlight which pixels in an image most strongly influence the model's prediction. Computed by calculating gradients of the model's output with respect to input pixels, saliency maps produce heatmaps where brighter regions indicate pixels that, when changed, would most significantly affect the prediction. This technique helps users understand which parts of an image the model is 'looking at' when making decisions.	Explainability Fairness
Concept Activation Vectors	Concept Activation Vectors (CAVs), also known as Testing with Concept Activation Vectors (TCAV), identify mathematical directions in neural network representation space that correspond to human-understandable concepts such as 'stripes', 'young', or 'medical equipment'. The technique works by finding linear directions that separate activations of concept examples from non-concept examples, then measuring how much these concept directions influence the model's predictions. This provides quantitative answers to questions like 'How much does the concept of youth affect this model's hiring decisions?' enabling systematic bias detection and model understanding.	Explainability Fairness Transparency
Differential Privacy	Differential privacy provides mathematically rigorous privacy protection by adding carefully calibrated random noise to data queries, statistical computations, or machine learning outputs. The technique works by ensuring that the presence or absence of any individual's data has minimal impact on the results - specifically, any query result should be nearly indistinguishable whether or not a particular person's data is included. This is achieved through controlled noise addition that scales with the query's sensitivity and a privacy budget (epsilon) that quantifies the privacy-utility trade-off. The smaller the epsilon, the more noise is added and the stronger the privacy guarantee, but at the cost of reduced accuracy.	Privacy Transparency Fairness
Attention Visualisation in Transformers	Attention Visualisation in Transformers analyses the multi-head self-attention mechanisms that enable transformers to process sequences by attending to different positions simultaneously. The technique visualises attention weights as heatmaps showing how strongly each token attends to every other token across different heads and layers. By examining these attention patterns, practitioners can understand how models like BERT, GPT, and T5 build contextual representations, identify which tokens influence predictions most strongly, and detect potential biases in how the model processes different types of input. This provides insights into positional encoding effects, head specialisation patterns, and the evolution of attention from local to global context across layers.	Explainability Fairness Transparency