Decoding Data Bias in Machine Learning: Key Takeaways from CSCI S-184, Harvard Summer School Class in the Trenches of AI Ethics
In machine learning and data science, ensuring fairness and avoiding bias is paramount. As we navigate the complexities of this field, the Harvard Summer School class on Data Science and AI Ethics, Regulations, and Laws has sparked valuable discussions on these topics. This post draws on one of those insightful dialogues to shed light on the nuanced aspects of data bias in machine learning.
What is Bias in Machine Learning?
In machine learning, bias is a model's inclination or predisposition towards certain outcomes or predictions, often due to the data used to train it. It's important to note that not all biases are harmful or unwanted - biases can merely reflect the non-uniform distribution often found in real-world datasets. The key is recognizing and understanding these biases, ensuring they do not introduce unwarranted prejudice or lead to incorrect conclusions.
Understanding the Role of Bias
Take, for instance, a case study discussed in the CSCI S-184 class. We examined a study aiming to identify the drivers of diabetes in a region where the average Body Mass Index (BMI) is greater than the national average. Here, a dataset with a higher proportion of individuals with a BMI above 25 would be representative of the population being studied, thus making it appropriate.
However, if the study aims to understand diabetes drivers more generally across populations with varying BMIs, then a dataset mostly comprised of individuals with high BMI could introduce bias. In this scenario, the study's results may only apply to populations with a higher average BMI, potentially overlooking factors more relevant in populations with lower BMIs.
Recommended by LinkedIn
Going Beyond Sex, Gender, and Race
Our class discussions have underlined that fairness in machine learning isn't only about demographic factors like sex, gender, and race. Any variable that can lead to unfair representation or treatment can be a source of bias. This includes age, BMI, income, education level, and more.
The goal should be to strive for a representative sample that captures the diversity and variation of the population you are studying. By doing so, we can ensure the model's predictions are accurate and fair across different groups of people.
Understanding and accounting for bias in machine learning is more than just a theoretical exercise. It is a significant contributor to the accuracy and fairness of our models. By considering bias in our datasets, we can create models that perform well while being fair and unbiased.
These rich discussions in the Harvard Summer School class underscore the importance of carefully considering the design and composition of your dataset when conducting a study. We move one step closer to achieving fairness in machine learning by considering these factors.
Missed out on this summer's exploration into AI ethics? Don't worry; the quest for unraveling the nuances of data bias in machine learning isn't over yet. CSCI S-184 is coming back in Spring 2024. Grab your spot and join us on this exciting journey to decode the ethics, regulations, and laws in Data Science and AI.
