Fairer non-negative matrix factorization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Growing need for fairness and bias mitigation in machine learning (ML) algorithms.
• Existing methods lack a one-size-fits-all solution, necessitating practical bias mitigation strategies.
• Non-negative matrix factorization (NMF) is a key technique for topic modeling and feature extraction.

Purpose of the Study:

• To adapt non-negative matrix factorization (NMF) for improved fairness in ML.
• To develop practical bias mitigation strategies for NMF.
• To explore the trade-offs between fairness and accuracy in ML models.

Main Methods:

• Modified the objective function of NMF using a min-max formulation.
• Derived two optimization methods: a multiplicative update rule and an alternating minimization scheme.
• Evaluated the approach using synthetic and real-world datasets.

Main Results:

• The min-max NMF formulation can sometimes improve fairness for specific population groups.
• The method's effectiveness is dependent on the specific application and data.
• Potential trade-off observed: improved fairness may sometimes increase individual error rates.

Conclusions:

• Fairness in ML is not a rigid definition and requires application-specific solutions.
• The proposed NMF modification offers a potential tool for enhancing fairness.
• Practitioners must carefully weigh fairness improvements against potential impacts on individual accuracy.