With the widespread use of artificial intelligence (AI) systems and applications in our everyday lives, accounting for fairness has gained significant importance in designing and engineering of such systems. AI systems can be used in many sensitive environments to make important and life-changing decisions; thus, it is crucial to ensure that these decisions do not reflect discriminatory behavior toward certain groups or populations. More recently some work has been developed in traditional machine learning and deep learning that address such challenges in different subdomains. With the commercialization of these systems, researchers are becoming more aware of the biases that these applications can contain and are attempting to address them.
In the industry, it has become very critical to create fair ML models in order to respect different groups in the sensitive features that are protected by the law and not to favorably select some groups against the others. Bias can show up in either dataset sampling or model performance against protected groups or individuals. Therefore, it is important in the industry to establish a bias analysis system to identify and mitigate the bias in both the dataset and model performance with respect to group and individual fairness.
There are several fairness libraries to achieve this job. In the industry, fairness libraries that are used in bias analysis must be created by well-known organizations. There are fairness libraries created by big companies such as Microsoft, IBM, and Google. The goal of this project is to compare the fairness libraries that can be used in the industry and work out a use-case using a published dataset.
Project Information Subsection
1. Surveying the basics of bias and fairness in machine learning. The students will learn the basics from the two review articles “A Survey on Bias and Fairness in Machine Learning” by NINAREH MEHRABI, FRED MORSTATTER, NRIPSUTA SAXENA, KRISTINA LERMAN, and ARAM GALSTYAN, and “An Introduction to Algorithmic Fairness” arXiv:2105.05595v1 [cs.CY] by Hilde J.P. Weerts.
2. Searching for possible fairness libraries that can be used in the industry. We will use three libraries created by big technology companies, so that they are trustable to be used in industry.
• Fairlearn (By Microsoft)
• AIF360 (By IBM)
• What-if-tool (By Google)
3. Selecting a published structured and unstructured dataset. The main goal of the project is to mitigate bias in the structured (tabular) dataset. If possible, we will extend our bias analysis to the unstructured data such as text and image.
• Tabular Dataset: TitanicSexism (fairness in ML), https://www.kaggle.com/code/garethjns/titanicsexism-fairness-in-ml/input
• Text Dataset: Fake and real news dataset, https://www.kaggle.com/datasets/clmentbisaillon/fake-and-real-news-dataset
• Imaged Dataset: UTKFace, https://www.kaggle.com/datasets/jangedoo/utkface-new
4. Discussing the possible mitigation algorithms that can be used. Mitigation algorithms should be implemented in pre-processing, in-processing, and post-processing. Below is an example of the mitigation algorithms that will be used.
• Fairlearn: ExponentiatedGradient, GridSearch, ThresholdOptimizer, CorrelationRemover, AdversarialFairnessClassifier, AdversarialFairnessRegressor.
• AIF360: preprocessing (Disparate Impact Remover, LFR, Optim Preproc, Reweighing), inprocessing (Adversarial Debiasing, ART Classifier, Gerry Fair Classifier, Meta Fair Classifier, Prejudice Remover, Exponentiated Gradient Reduction, GridSearch Reduction), postprocessing (Calibrated EqOdds Postprocessing, EqOdds Postprocessing, Reject Option Classification).
• What-If-Tool: It is still under study
5. Discussing the results and summarizing the comparison among the libraries. In the discussion, we will compare the performance of the mitigation algorithms in different stage of the ML life cycle such as preprocessing, inprocessing, and postprocessing.
{Empty}
PI has an undergraduate student they would like to work with them on this project.
{Empty}
{Empty}
{Empty}
{Empty}
6127 Galleon Dr Mechanicsburg, Pennsylvania. 17050
Milestone Title: Survey basics Milestone Description: Surveying the basics of bias and fairness in machine learning. The students will learn the basics from the two review articles “A Survey on Bias and Fairness in Machine Learning” by NINAREH MEHRABI, FRED MORSTATTER, NRIPSUTA SAXENA, KRISTINA LERMAN, and ARAM GALSTYAN, and “An Introduction to Algorithmic Fairness” arXiv:2105.05595v1 [cs.CY] by Hilde J.P. Weerts. Completion Date Goal: 2023-10-26 Actual Completion Date: 2023-11-13
Milestone Title: Select libraries Milestone Description: Choosing the proper fairness metrics to identify the bias. Below is an example of the metrics that will be used in each library.
• Fairlearn: Demographic parity, Equalized odds, Equal opportunity
• AIF360: Dataset Metric, Binary Label Dataset Metric, Classification Metric, Sample Distortion Metric, MDSS Classification Metric.
• What-If-Tool: It is still under study Completion Date Goal: 2023-11-02 Actual Completion Date: 2023-11-13
Milestone Title: Select dataset Milestone Description: Selecting a published structured. The main goal of the project is to identify bias in the structured (tabular) dataset. If possible, we will extend our bias analysis to the unstructured data such as text and image.
• Tabular Dataset: TitanicSexism (fairness in ML), https://www.kaggle.com/code/garethjns/titanicsexism-fairness-in-ml/input
• Text Dataset: Fake and real news dataset, https://www.kaggle.com/datasets/clmentbisaillon/fake-and-real-news-dataset
• Imaged Dataset: UTKFace, https://www.kaggle.com/datasets/jangedoo/utkface-new Completion Date Goal: 2023-11-09 Actual Completion Date: 2023-11-13
Milestone Title: Choose fairness metrics Milestone Description: Choosing the proper fairness metrics to identify the bias. Below is an example of the metrics that will be used in each library.
• Fairlearn: ExponentiatedGradient, GridSearch, ThresholdOptimizer, CorrelationRemover, AdversarialFairnessClassifier, AdversarialFairnessRegressor.
• AIF360: preprocessing (Disparate Impact Remover, LFR, Optim Preproc, Reweighing), inprocessing (Adversarial Debiasing, ART Classifier, Gerry Fair Classifier, Meta Fair Classifier, Prejudice Remover, Exponentiated Gradient Reduction, GridSearch Reduction), postprocessing (Calibrated EqOdds Postprocessing, EqOdds Postprocessing, Reject Option Classification).
• What-If-Tool: It is still under study
Completion Date Goal: 2023-12-07
Milestone Title: Identify bias in dataset Milestone Description: Employ the selected libraries on the dataset and identify bias Completion Date Goal: 2023-12-22
Milestone Title: Discussing results and summarizing comparison Milestone Description: Discussing the results and summarizing the comparison among the libraries. In the discussion, we will compare the performance of the mitigation algorithms in different stage of the ML life cycle such as preprocessing, inprocessing, and postprocessing.
Completion Date Goal: 2024-01-18
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
Mentor is needed - skills in machine learning and bias identification needed.
Final Report
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
{Empty}
Feedback from the PI: "Process was very smooth and professional"
Feedback from the student facilitator: "The project went better than he was expecting and he gained a lot of experience and learning. Didn't previously have knowledge or awareness of fairness when it came to machine learning."
Feedback from mentor: "Very relevant research, particularly for industry applications. Undergraduates doing hands-on research is a great experience. Structure and detailed milestones were very helpful."