Abstract
Completed during the fall of 2022, the objective of my capstone project was to explore ethics in data science as well as to build a model to determine utilization in Medical Panel Expenditure Survey (MEPS) data. Utilizing IBM's AI Fairness 360 toolkit, I was able to practice pre-processing, in-processing, and post-processing techniques to mitigate bias along with determining relevant fairness metrics to detect bias. This capstone project was mentored by industry partners from Deloitte providing helpful feedback in the context of ethical industry practices.
The original data used was from the MEPS survey conducted in 2014/15. Within the dataset, an exploratory data analysis was conducted eventually narrowing down the data to 43 relevant features regarding a patient's health. These features consisted of metrics such as identifiers (race, age, sex, etc.) as well as various categorical variables regarding a patient's health characteristics (high cholesterol, history of heart attack, etc.). On top of the EDA, a correlation analysis was conducted to determine how the features correlate to one another.
The main purpose of this project concerned how 'RACE' which is considered a sensitive attribute has on a 'UTILIZATION' score. The UTILIZATION score is defined as the total number of trips requiring some sort of medical care by summing up the following features: OBTOTV15(16), the number of office-based visits; OPTOTV15(16), the number of outpatient visits; ERTOT15(16), the number of ER visits; IPNGTD15(16), the number of inpatient nights, and + HHTOTD16, the number of home health visits.
A model was then built with a classification task to predict whether a person would have 'high' utilization (defined as UTILIZATION >= 10, roughly the average utilization for the considered population).
The other main goal of the project was to understand how bias in models can occur and methods to detect and mitigate bias. On this front, IBM's open-source package AI Fairness 360 was used to help accomplish these tasks. In essence what AI Fairness 360 (AIF 360) provides is a comprehensive set of metrics for datasets and models to test for biases, explanations for these metrics, and algorithms to mitigate bias in datasets and models.
The next step of building the actual classifier models was then broken down into 2 steps:
- Training models without de-biasing, using IBM's tutorial
- Training models without de-biasing, using our own model development techniques including Feature Selection, Encoding, Binning Features, and other items
In the original IBM tutorial, two models were built (Logistic Regression and Random Forest) and were evaluated using a balanced accuracy score. For the capstone project, my group and I decided to also build a Logistic Regression and Random Forest model(s) but experimented with changing the number of leaves/trees for the random forest model and using an L2 loss penalty term on the regression model. Furthermore, we believed that instead of measuring performance with balanced accuracy, measuring with disparate impact (proportion of the unprivileged group that received the positive outcome divided by the proportion of the privileged group that received the positive outcome) would better suit our purpose as it is metric which measures demographic parity. For our project, the intended goal was indeed to mitigate bias in UTILIZATION among demographic groups.
After building the model(s) is when we began to experiment with the de-biasing processing techniques. We first experimented with the preprocessing technique of reweighting which reweights class labels. In essence, what this does is punish the unprivileged class less for being classified with a favorable outcome and punish the privileged class less for being classified with an unfavorable outcome. Performing this technique saw an improvement in fairness metrics, but the improvement was marginal and did not achieve true fairness (disparate impact of less than 0.2).
The next technique was experimented with was the in-processing technique of prejudice remover which adds a discrimination-aware regularization term to the learning objective. In essence, it allows the classifier to be independent of any sensitive features/information. After performing this technique, we achieved significantly positive results with true "fairness" achieved in terms of disparate impact. The one tradeoff that was observed however is that by doing so, accuracy was lower in exchange for greater "fairness".
For the last technique, we experimented with the post-processing technique of the reject-option classifier. In essence, a reject option classifier takes place after the model determines its results and fits a grid of classification thresholds where "option rejection" occurs. What happens is it will give favorable outcomes to unprivileged groups and unfavorable outcomes to privileged groups in a confidence band around the decision boundary with the highest uncertainty. In doing so, we achieve the greatest fairness and accuracy out of all techniques.
For the final part of our project, we came to the conclusion that the best technique was the prejudice remover as while the reject option classifier offered the best accuracy along with the greatest fairness when considering real-world use, a reject option classifier would be ethically illegal due to in essence practicing reverse discrimination.
Through this project, my team and I learned how to use IBM's AI Fairness 360 toolkit, how to statistically detect bias in models, and practice methods for achieving statistical fairness. Overall, this capstone project pushed us to utilize the techniques we have gathered through our time at UC San Diego's data science program and apply them to a real-world use case where many difficult decisions had to be made. Furthermore, we were very fortunate to have industry mentors from Deloitte guide us through this project and help answer the questions we had. Overall, my team and I felt very successful in our work and believe that our project/investigation was completed in a professional, thorough, and ethical manner.