Disparity mitigation in machine learning-based predictions for distinct classes of data using derived indiscernibility constraints during neural network training
A system and method includes generating approximate distributions for distinct classes of data samples; computing a first partial Jensen-Shannon (JS) divergence and a second partial JS divergence based on the approximate distribution of the disparity affected class of data samples with reference to the approximate distribution of the control class of data samples; computing a disparity divergence based on the first partial JS divergence and the second partial JS divergence; generating a distribution-matching term based on the disparity divergence, wherein the distribution-matching term mitigates an inferential disparity between the control class of data samples and the disparity affected class of data samples during a training of an unconstrained artificial neural network; constructing a disparity-constrained loss function based on augmenting a target loss function with the distribution-matching term; and transforming the unconstrained ANN to a disparity-constrained ANN based on a training of the unconstrained ANN using the disparity-constrained loss function.
1. A computer-implemented method for mitigating disparities of inferential outputs of a machine learning model, the method comprising:
generating an indiscernibility constraint based on a distribution divergence metric value computed between a distribution associated with a non-disparity affected class of data samples and a distribution associated with a disparity affected class of data samples, wherein the indiscernibility constraint mitigates a machine learning-based inferential disparity between the non-disparity affected class of data samples and the disparity affected class of data samples during a training of a target machine learning model;
configuring a disparity-mitigating loss function based on augmenting a gradient descent algorithm with the indiscernibility constraint; and
training the target machine learning model using a training corpus of labeled data samples, wherein the training includes performing backpropagation using the disparity-mitigating loss function.
2. The method according to claim 1 , wherein
the disparity-mitigating loss function is defined as follows:
l
(
𝒜
,
ℬ
,
X
,
Y
)
=
λ
∑
i
ξ
(
a
i
,
b
i
)
︸
Reduce
error
+
(
1
-
λ
)
JS
(
X
,
Y
,
m
)
︸
reduce
disparity
where (A, B)=| (a i , b i )∈<R N , R> comprises a sequence of data samples and target values corresponding to the sequence of data samples, where | relates to a target loss function, where X, Y⊂A indicates that a first set of data samples X and a second set of data samples Y are subsets of a corpus of data samples A, where | ∈[0, 1], where m>0.
3. The method according to claim 2 , wherein:
the target loss function guides the gradient descent algorithm;
the first set of data samples comprises the non-disparity affected class of data samples; and
the second set of data samples comprises the disparity affected class of data samples.
4. The method according to claim 1 , wherein
the indiscernibility constraint, when applied in the training of the target machine learning model, reduces a Jensen-Shannon divergence between a distribution of a first set of machine learning inferences based on a first class of data and a distribution of a second set of machine learning inferences based on a second class of data.
5. The method according to claim 1 , wherein
the indiscernibility constraint, when applied in the training of the target machine learning model, causes a distribution of a first set of machine learning inferences based on a first class of data to match a distribution of a second set of machine learning inferences based on a second class of data.
6. The method according to claim 1 , wherein
the indiscernibility constraint, when applied in the training of the target machine learning model, causes a distribution of a first set of machine learning inferences based on a first class of data to align with a distribution of a second set of machine learning inferences based on a second class of data.
7. The method according to claim 1 , wherein
the backpropagation of the training of the target machine learning model, when applying the disparity-mitigating loss function, computes coefficients for learnable variables of the target machine learning model that simultaneously minimizes an accuracy error of the target machine learning model while minimizing a disparity distance between distinct distributions of inferential outputs of the target machine learning model.
8. The method according to claim 1 , wherein
generating the indiscernibility constraint includes computing a first partial Jensen-Shannon (JS) divergence based on an approximate distribution of the non-disparity affected class of data samples with reference to an approximate distribution of the disparity affected class of data samples.
9. The method according to claim 8 , wherein
generating the indiscernibility constraint includes computing a second partial JS divergence based on the approximate distribution of the disparity affected class of data samples with reference to the approximate distribution of the non-disparity affected class of data samples.
10. The method according to claim 9 , wherein
generating the indiscernibility constraint includes computing a JS divergence distance based on the first partial JS divergence and the second partial JS divergence, and
the JS divergence distance relates to a distance between the approximate distribution of the non-disparity affected class of data samples and the approximate distribution of the disparity affected class of data samples that indicates a likely disparity between machine learning inferences associated with member samples of the disparity affected class and the non-disparity affected class.
11. The method according to claim 1 , further comprising:
computing a JS divergence distance based on machine learning inferences of the target machine learning model, as trained, for each of a first class of data samples and a second class of data samples; and
terminating a continuation of the training of the target machine learning model when the JS divergence distance satisfies a non-disparity threshold comprising a maximum JS divergence distance value.
12. The method according to claim 1 , further comprising:
computing a distinct indiscernibility constraint for each of a plurality of distinct pairings of a given non-disparity affected class of data samples and a given disparity-affected class of data samples.
13. The method according to claim 1 , further comprising:
sourcing a corpus of data samples comprising at least (a) a non-disparity affected sub-corpus of the non-disparity affected class of data samples and (b) a disparity-affected sub-corpus of the disparity affected class of data samples, wherein sourcing the corpus of data samples includes sourcing data samples from within a defined geographical region, wherein the corpus of data samples comprises a geographic region-specific corpus of data samples, and
generating the indiscernibility constraint is based on the geographic region-specific corpus of data samples.
14. A method of encouraging inferential fairness of an artificial neural network, the method comprising:
computing a disparity divergence based on an approximate distribution of a non-disparity affected class of data samples and an approximate distribution of a disparity affected class of data samples;
generating a distribution-matching term based on the disparity divergence, wherein the distribution-matching term mitigates an inferential disparity between artificial neural network inferences for the non-disparity affected class of data samples and the disparity affected class of data samples during a training of an artificial neural network;
constructing a disparity-constrained loss function based on integrating the distribution-matching term with a loss function; and
training the artificial neural network using a training corpus of labeled data samples, wherein the training includes performing backpropagation using the disparity-constrained loss function.
15. The method according to claim 14 , wherein
the distribution-matching term encourages a distribution of a first set of artificial neural network inferences based on a first class of data toward a distribution of a second set of artificial neural network inferences based on a second class of data.
16. A method comprising:
computing a disparity divergence based on an estimated distribution of a disparity affected class of data samples and an estimated distribution of a control class of data samples;
generating a distribution-matching term based on the disparity divergence, wherein the distribution-matching term mitigates an inferential disparity between the control class of data samples and the disparity affected class of data samples during a training of an unconstrained artificial neural network;
constructing a disparity-constrained loss function based on including the distribution-matching term with a target loss function; and
transforming the unconstrained artificial neural network to a disparity-constrained artificial neural network based on a training of the unconstrained artificial neural network using the disparity-constrained loss function.
17. The method according to claim 16 , wherein
the distribution-matching term, when applied in the training of the unconstrained artificial neural network, causes a distribution of a first set of artificial neural network inferences based on a first class of data to align with a distribution of a second set of artificial neural network inferences based on a second class of data.
18. The method according to claim 16 , wherein
training the unconstrained artificial neural network further includes performing backpropagation using the disparity-constrained loss function, wherein the backpropagation, when applying the disparity-constrained loss function, computes coefficients for learnable variables of the disparity-constrained artificial neural network that simultaneously minimizes an accuracy error of the disparity-constrained artificial neural network while minimizing a disparity distance between distinct distributions of inferential outputs of the disparity-constrained artificial neural network.