Podcast
Fairness in AI/ML Engineering: Interpretability, Metrics and Sociotechnical Design
Open original DataTalks.Club episode
Fairness in AI/ML Engineering: Interpretability, Metrics and Sociotechnical Design
Original Episode
Use these links for the canonical episode and media sources.
- Open the original DataTalks.Club podcast page
- Watch on YouTube
- Listen on Spotify
- Listen on Apple Podcasts
Episode Overview
How do you reduce bias in credit scoring models without sacrificing explainability? In this episode, Tamara Atanasoska — an open source software engineer at :probabl.., Fairlearn maintainer, and contributor to scikit-learn and skops with a background in software engineering and computational linguistics — walks through practical approaches to fairness in AI. We dig into a real credit scoring use case, empirical findings on gender disparities, and the societal harms of biased models such as debt and repossession.
People
Use these links to connect the episode to guest notes.
Chapter Summary
Use these checkpoints to decide whether to open the source transcript.
- 0:00 - Podcast Introduction & Episode Overview
- 2:31 - Guest Introduction: Tamara’s Open-Source Roles (Fairlearn, scikit-learn,
- 3:18 - Career Overview: Software Engineering to Computational Linguistics
- 4:37 - Music Tech Experience: Ableton and Push 2 Instrument Design
- 6:41 - Device Architecture: Laptop Computation vs Standalone Hardware
- 8:56 - Transition to NLP & AI: Academic Interests and Motivation
- 10:04 - Cognitive Systems Studies: Language, Neuroscience, and ML
- 11:50 - Research Path: Returning to Study and New Projects
- 12:41 - Music as Hobby: Balancing Creative Work and Research
- 13:44 - Sociotechnical Framing: Modeling Language in Social Context
- 14:52 - Fairness in AI: Credit Scoring Use Case and Real-World Impact
- 15:10 - Empirical Findings: Gender Disparities in Credit Models (Fairlearn Study)
- 18:14 - Societal Harms: Debt, Repossession, and Downstream Consequences
- 21:31 - Fairlearn Tools: Group Fairness, Visualization, and Mitigation Methods
- 24:04 - Sensitive Group Selection: Domain-Specific Decisions in Credit Models
- 26:21 - Limits of Automation: Human Judgment in Fairness Assessments
- 28:52 - Metric Tradeoffs: False Positives vs False Negatives and Demographic Parity
- 31:33 - Organizational Responsibility: Who Decides Fairness Tradeoffs?
- 33:11 - Practitioner Education: Frameworks, Ambiguity, and Learning Objectives
- 35:23 - Moderation Case Study: Cross-Functional Teams and Domain Expertise
- 37:13 - Human-in-the-Loop: Essential Component for Fair AI Systems
- 39:18 - Joining Probable: From Open-Source Contributions to a Role
- 40:57 - Probable Work: Explainability, Language Models, and Library Integration
- 42:54 - Interpretability Tools: Inspection Package and Partial Dependence
- 44:54 - Cross-Library Compatibility: Fairlearn, scikit-learn, and Estimator APIs
- 46:20 - Scopes Library: Secure Model Persistence and Hugging Face Integration
- 47:16 - Serialization Risks: Pickle Vulnerabilities and Secure Deserialization
- 50:54 - Community Involvement: PyLadies, Sprints, and Fairlearn Events
- 52:10 - Contributing to Fairlearn: Discord, Good-First Issues, and Sprints
- 55:41 - Development Ethos: Testing, Refactoring, and Custom Estimators
- 56:37 - Project Updates: Upcoming Fairlearn Release and Maintainer Notes
- 57:22 - Practical Quirk: Tokenization Issues Breaking “Fairlearn” in Transcripts
- 58:14 - Closing Remarks, Contact Info, and Final Thoughts