Certifying Some Distributional Fairness with Subpopulation Decomposition

Certifying Some Distributional Fairness with Subpopulation Decomposition

Verify some distributional fairness using subgroup decomposition

Background : Extensive efforts have been made to understand and improve machine learning model fairness based on different fairness metrics, especially in high-risk domains such as health insurance, education, hiring decisions, etc.

Problems and flaws : The performance of end-to-end machine learning models lacks verifiable fairness .

Method and Model :

On a given data distribution, validation fairness of a trained ML model is formulated as an optimization problem based on model performance loss bounds on the distribution of fairness constraints within a bounded distribution distance of the training data distribution.

Contribute :

  1. A general fairness verification framework is proposed, and instantiated verification with sensitive offsets and general offsets .
  2. Decompose the source data distribution into analyzable subgroup distributions, solve the subproblems by proving their convexity, and thus solve the model optimization problem.
  3. Experiments prove that the verifiability of the model is strict in the case of sensitive offsets, and it is strict in the case of general offsets.extraordinaryof.
  4. The framework can flexibly integrate additionalunbiased constraint, and the result will be more stringent.
  5. The proposed verifiable fairness bounds are compared with existingAdaptive Distribution Robust BoundsA comparison shows that the former is more stringent.

1 Introduction

Flaws in previous work :

  1. Previous work mainly focuses on regularization training, disentanglement, duality, low-order matrix decomposition, distribution alignment and other methods to improve the fairness of ML.

  2. Some work on verifiable fairness characterization has been done on ML, but there is a problem: training an end-to-end model on a randomly given data distribution, this model lacks verifiable fairness in predicting results.

  3. The ML model that the existing literature on fairness focuses on is to train the model on a (non-)balanced data distribution, and evaluate the performance of the model through the existing fairness evaluation method in the measurable target domain, so the fairness The evaluation depends only on the choice of evaluation method, and does not take into account the validity of the method.

Verifiable Fairness : On a fairness-constrained test distribution Q, where Q is within a bounded distance from the training distribution P, define verifiable fairness as the worst-case upper bound on the model's prediction loss .

Base Rate Conditionsas a fairness constraint on the test distribution Q.

sensitive shifting Sensitive shift : The cascade distribution of sensitive attributes and labels can change.

general shifting : everything including the conditional distribution of non-sensitive attributes can be changed.

Group Fairness : Measuring the independence between sensitive features and model predictions. Separation [separation] indicates that given the target label, the sensitive feature is statistically independent of the model prediction. Sufficiency [sufficiency] indicates that a given model prediction, sensitive features are statistically independent of the target label. That is, group fairness requires sensitive attributes to be independent of target labels and model predictions.

Individual fairness : similar input features will produce similar output results.

This paper differs from previous work :

  1. Validation fairness considers the performance of end-to-end ML models, not the level of representation learning.
  2. Fairness is defined and verified based on a distribution of fairness constraints.
  3. For any black-box model trained on a given random data distribution, verifiable fairness can be computed.

Question 1: How does the benchmark interest rate conditional constraint encode the fairness constraint distribution?

2. Verifiable fairness based on fair constraint distribution

Definition 1 Benchmark interest rate : Given a distribution PP supported on X*YP , sensitive attribute valuesss relative to labelyyThe base rate of y is:SSS is the sensitive attribute, [S] is the set of values ​​that the sensitive attribute can take,sss is a sensitive attribute value of a sensitive attribute. YYY is the prediction result of the model,yyy is the sample label,XXX is the sample feature. The benchmark interest rate of a test sample is [sensitive attribute characteristicX s X_sXsThe value is sss , the predicted result isyyProbability of y ]:

image-20221114181417296

Definition 2. Fair benchmark interest rate distribution : if and only if in a distribution generated by the benchmark interest rate, for any two samples iii andjjj , both have the same predicted labelyyy , and both have the sensitive attributeSSFor a certain attribute value in S , their corresponding benchmark interest rates are equal, and this distribution is called a fair benchmark interest rate distribution:

image-20221114181507813

Demographic Parity : Group fairness evaluation index

image-20221117141519913

2.1 Verifiable fairness

Data generation model : X o X_oXoRepresents non-sensitive attribute features, X s X_sXsRepresents sensitive attribute characteristics, YYY represents the sample label.

image-20221115225158524

Verification fairness with general offset : P \mathcal PP is the training set distribution,ρ \rhoρ is the test set distributionQ \mathcal QQ andP \mathcal PThe distribution distance bound between P. For all generated and training set distributions at distribution distanceρ \rhoThe test set distribution within the range of ρ , the fairness proof value is: the maximum value of the upper bound of the loss value in all test set distributions.

image-20221114181040565

In real scenarios, the model training set is always limited by data management and data collection, so there is always inherent unfairness in the trained model. Assuming that the test set we construct is ideally fair, we hope that our model will not encode the bias generated during training when testing, so the performance of the model on the fair constraint distribution indicates its inherent unfairness.

Verifiable Fairness for Sensitive Offsets : In order to avoid the inherent unfairness of the model during training, a new constraint is added under the verification fairness of general offsets. P s , y P_{s,y}Pyes , andFrom Q s , y Q_{s,y}Qyes , andis PPP andQQThe subgroup of Q [by the sensitive attribute sssyyy division].

image-20221114181115540

To constrain the test set distribution QQQ is not too skewed towards sensitive attributesX s X_sXs, so an additional constraint term is added to the sensitive offset:

image-20221117150943254

l o s s loss l oss in each group and each category:

image-20221117152013594

The above loss can be transformed into ε-DP and ε-EO :

image-20221117152440782

Group parity (Demographic parity, DP): The difference in the predicted probability of predicting two different groups as positive classes.

Equalized odds: The difference between false-positive rates between groups or the difference between true-positive rates between groups. The smaller the difference, the fairer the model.

3. Fairness certification framework

3.1 Subgroup Decomposition [Core]

Hellinger distance : Measures the distance between two distributions. The value range is [0, 1], the larger the more relevant

image-20221117153124935

Overall optimization problem :

image-20221118105031278

image-20221118105043370

image-20221117153902026

image-20221118104406155

Subgroup optimization problem :

image-20221118104514617

image-20221118111433319

3.2 Verifiable Fairness with Sensitive Offset

image-20221118105832265

image-20221118111316377

image-20221118111326577

image-20221118111759859

image-20221118111902585

3.3 Verifiable Fairness with General Offset

image-20221118112131894

image-20221118112152904

Loss calculation after subgroup decomposition :

image-20221118113056217

image-20221118140325819

image-20221118140340288

image-20221118140414378

appendix

Scalar: A scalar is a single number (integer or real), unlike most other objects studied in linear algebra (usually arrays of numbers). Scalars are usually denoted by italic lowercase letters, for example: x \mathit xx , a scalar is equivalent to the one defined in Python

x = 1

Vector (vector): A vector represents a set of ordered numbers. We can find each individual number through the index in the order. Vectors are usually represented by bold lowercase letters, for example: x \bf xx , each element in the vector is a scalar, the i-th element in the vector is represented by $x_i$, and the vector is equivalent to a one-dimensional array in Python

import numpy as np
#行向量
a = np.array([1,2,3,4])

Matrix (matrix): A matrix is ​​a two-dimensional array, each element of which is determined by two indices ( A i , j A_{i,j}Ai,j), the matrix is ​​usually represented by bold italic capital letters, for example: $ \boldsymbol X$. We can think of the matrix as a two-dimensional data table, each row of the matrix represents an object, and each column represents a feature. Defined in Python as

import numpy as np
#矩阵
a = np.array([[1,2,3],[4,5,6],[7,8,9]])

Tensor: An array of more than two dimensions. Generally speaking, the elements in an array are distributed in a regular grid of several-dimensional coordinates, which is called a tensor. If a tensor is a three-dimensional array, then we need three indices to determine the position of the elements ( A i , j , k A_{i,j,k}Ai,j,k), tensors are usually represented by bold capital letters, for example: X \bf XX

mport numpy as np
#张量
a = np.array([[[1,2],[3,4]],[[5,6],[7,8]]])

bound constraint to find the range

worst case boundary

How to calculate the r of each test sample, when r<bound is certified

Within the bound range, the prediction result must be greater than or equal to the worst case, so it has the nature of verification

Verifiable properties: In the prediction results, how many samples are in line with the user-specified bound prediction R.

data distribution

download

Guess you like

Origin blog.csdn.net/qq_45724216/article/details/129865478