MORE | Spring 2023

Mitigating Label Bias Through Probabilistic Modeling With Latent Variable

Data icon, disabled. Four grey bars arranged like a vertical bar chart.

Supervised learning tasks rely on training data to make predictions and often assume them to be balanced and representative of real-world data. This proposal studies the effectiveness of notions of fairness like Equalized Odds (EO) and Equal Opportunity (EOP) in presence of label bias. We aim to discover hidden fair labels through probabilistic modeling and utilize them to enforce notions of fairness. We evaluate our approach against a synthetic dataset and demonstrate fairer decision-making.

Student researcher

Saurav Anchlia

Computer science

Hometown: Burdwan, West Bengal, India

Graduation date: Spring 2024