Paired-consistency-based model-agnostic approach to fairness in machine learning models

1. A computer implemented method for analyzing a machine learning model, said method being performed on a computing device, said method comprising:

inputting a dataset for use with the machine learning model, the input dataset comprising one or more features associated with a protected feature of users identified within the dataset;

generating by a domain expert a plurality of consistency pairs based on at least one feature from the one or more features within the input dataset, each consistency pair containing different examples of the protected feature but warranting a similar response by the machine learning model; and

analyzing the machine learning model using the plurality of consistency pairs and the input dataset.

2. The method of claim 1 , wherein said analyzing step comprises performing at least one of a model fairness evaluation process or a model fairness regulation process on the machine learning model.

3. The method of claim 2 , wherein performing the model fairness evaluation process comprises:

inputting the plurality of consistency pairs and the input dataset into the machine learning model;

determining, based on an output of the machine learning model, a precision score, a recall score and a paired-consistency score for the model; and

determining a fairness of the machine learning model based on the determined precision score, recall score and paired-consistency score.

4. The method of claim 2 , wherein performing the model fairness evaluation process comprises:

inputting the plurality of consistency pairs and the input dataset into the machine learning model;

determining, based on an output of the machine learning model, a precision score, a recall score and a paired-consistency score for the model;

determining a harmonic mean of the precision score, recall score and paired-consistency score; and

determining the fairness of the machine learning model based on the harmonic mean of the precision score, recall score and paired-consistency score.

5. The method of claim 2 , wherein performing model fairness regulation process comprises:

inputting a paired-consistency score for the machine learning model into a loss function of the model; and

training the machine learning model with training data comprising a subset of the consistency pairs and the input dataset.

6. The method of claim 5 , wherein training the machine learning model comprises using a tree-based training process and the training comprises:

adding the paired-consistency score as an extension to a gini index used in to tree create a tree associated with the machine learning model; and

maximizing a number of consistency pairs to go in a same direction in the tree.

7. The method of claim 5 , wherein training the machine learning model comprises using the paired-consistency score in a logistic regression-based training process.

8. The method of claim 1 further comprising:

generating by the domain expert a weight value for each of the plurality of consistency pairs; and

analyzing the machine learning model using the weighted plurality of consistency pairs and the input dataset.

9. A system for analyzing a machine learning model, said system comprising:

a first computing device connected to a second computing device through a network connection, the first computing device configured to:

input a dataset for use with the machine learning model, the input dataset comprising one or more features associated with a protected feature of users identified within the dataset;

generate by a domain expert a plurality of consistency pairs based on at least one feature from the one or more features within the input dataset, each consistency pair containing different examples of the protected feature but warranting a similar response by the machine learning model; and

analyze the machine learning model using the plurality of consistency pairs and the input dataset.

10. The system of claim 9 , wherein said computing device analyzes the machine learning model by performing at least one of a model fairness evaluation process or a model fairness regulation process on the machine learning model.

11. The system of claim 10 , wherein performing the model fairness evaluation process comprises:

inputting the plurality of consistency pairs and the input dataset into the machine learning model;

determining, based on an output of the machine learning model, a precision score, a recall score and a paired-consistency score for the model; and

determining a fairness of the machine learning model based on the determined precision score, recall score and paired-consistency score.

12. The system of claim 10 , wherein performing the model fairness evaluation process comprises:

inputting the plurality of consistency pairs and the input dataset into the machine learning model;

determining, based on an output of the machine learning model, a precision score, a recall score and a paired-consistency score for the model;

determining a harmonic mean of the precision score, recall score and paired-consistency score; and

determining the fairness of the machine learning model based on the harmonic mean of the precision score, recall score and paired-consistency score.

13. The system of claim 10 , wherein performing the model fairness regulation process comprises:

inputting a paired-consistency score for the machine learning model into a loss function of the model; and

training the machine learning model with training data comprising a subset of the consistency pairs and the input dataset.

14. The system of claim 13 , wherein training the machine learning model comprises using a tree-based training process and the computing device performs the training by:

adding the paired-consistency score as an extension to a gini index used in to tree create a tree associated with the machine learning model; and

maximizing a number of consistency pairs to go in a same direction in the tree.

15. The system of claim 13 , wherein training the machine learning model comprises using the paired-consistency score in a logistic regression-based training process.

16. The system of claim 10 wherein the computing device is further configured to:

generate by the domain expert a weight value for each of the plurality of consistency pairs; and

analyze the machine learning model using the weighted plurality of consistency pairs and the input dataset.