Fairness in AI-Powered Employee Promotion Systems

Project Overview

Problem Statement

• AI is increasingly used to assist promotion decisions.
• These systems often learn from biased historical data.
• This often leads to bias, especially taking sensitive attributes (like gender) into account, causing bias situations like gender bias — promoting men more often than women.
• We're studying whether AI models treat employees across all subgroups fairly in promotion outcomes.

Research Questions

1. Should we automate decisions that shape people’s futures — and if so, how can we do it fairly?
2. How can we ensure those automated systems make decisions that are trustable enough?

Models in Use

The Baseline Model - MLP Classifier

The K-Means Method

An Ordinary Model handling Out-Of-Distribution (OOD) Scenarios

A Cluster-based Model handling OOD Scenarios

Explainability

• We use LIME as the explainer.

Here's a comparison of the top 5 positively and negatively contributed features across all models: