Table of Contents

Interface IBiasDetector<T>

Namespace
AiDotNet.Interfaces
Assembly
AiDotNet.dll

Defines an interface for detecting bias in machine learning model predictions.

public interface IBiasDetector<T>

Type Parameters

T

The numeric type used for calculations (e.g., double, float).

Remarks

For Beginners: This interface helps identify unfair treatment in machine learning models.

In machine learning, bias occurs when a model treats different groups of people unfairly. For example, a loan approval model might unfairly reject applications from certain demographic groups.

This interface provides methods to:

  • Detect when a model is biased against certain groups
  • Measure how severe the bias is
  • Compare different models to find which one is most fair

The bias score measures how unfair the model is. Important points:

  • Lower bias scores usually indicate fairer models
  • The score helps us identify which groups are being treated unfairly
  • We can use these measurements to improve our models and make them more equitable

Properties

IsLowerBiasBetter

Indicates whether lower bias scores represent better (fairer) models.

bool IsLowerBiasBetter { get; }

Property Value

bool

Remarks

For Beginners: This property tells you whether smaller numbers mean fairer models.

For bias detection, this is typically true - lower bias scores mean the model is treating all groups more equally. A bias score of 0 would mean perfect fairness.

This property returns:

  • true: if lower scores mean fairer models (typical for bias metrics)
  • false: if higher scores mean fairer models (less common)

Knowing this helps you correctly interpret and compare bias scores across different models.

Methods

DetectBias(Vector<T>, Vector<T>, Vector<T>?)

Detects bias in model predictions by analyzing predictions across different groups.

BiasDetectionResult<T> DetectBias(Vector<T> predictions, Vector<T> sensitiveFeature, Vector<T>? actualLabels = null)

Parameters

predictions Vector<T>

The model's predictions.

sensitiveFeature Vector<T>

The sensitive feature (e.g., race, gender) used to identify groups.

actualLabels Vector<T>

Optional actual labels for computing additional bias metrics.

Returns

BiasDetectionResult<T>

A result object containing bias detection metrics and analysis.

Remarks

For Beginners: This method checks if your model treats different groups fairly.

The input parameters:

  • predictions: The outputs your model produced
  • sensitiveFeature: Information about which group each person belongs to (like age, gender, etc.)
  • actualLabels: The correct answers (optional, but provides more detailed analysis if available)

The method analyzes whether your model gives similar results to all groups or if it favors some groups over others. For example, it might detect if a hiring model recommends hiring men more often than equally qualified women.

The result includes:

  • Whether bias was detected (true/false)
  • How severe the bias is
  • Which groups are affected
  • Specific metrics measuring different types of unfairness

IsBetterBiasScore(T, T)

Compares two bias scores and determines if the current score represents better (fairer) performance.

bool IsBetterBiasScore(T currentBias, T bestBias)

Parameters

currentBias T

The bias score to evaluate.

bestBias T

The best (fairest) bias score found so far.

Returns

bool

True if the current bias score is better (fairer) than the best bias score; otherwise, false.

Remarks

For Beginners: This method compares two bias scores and tells you which model is fairer.

When comparing models, you need to know which one is more fair. This method handles the comparison logic for you, taking into account whether higher or lower scores indicate better fairness.

The input parameters:

  • currentBias: The bias score of a model you want to evaluate
  • bestBias: The best (fairest) score you've seen so far

The method returns:

  • true: if the current model is fairer than the best model so far
  • false: if the current model is less fair than or equal to the best model

This is particularly useful when testing many different models to find the fairest one.

Edit this page