Ramon van den Akker. Fairness of machine learning models an overview and practical considerations
0 likes134 views
This document discusses fairness in machine learning models. It begins with motivating examples of algorithms that were found to be biased, such as a recidivism prediction tool that was biased against black individuals. It then covers operationalizing fairness through frameworks like transparency and explainability. Finally, it discusses approaches for achieving fairness by design, such as preprocessing the data, adding randomness to predictions, or tailoring new algorithms with fairness constraints. The author notes there are inherent tradeoffs between performance and fairness that require difficult choices.
![new tailor made algorithms / model
common structure:
max “performance” subject to “fairness
constraints”
max [“performance” + 𝛼“fairness”]
in general there will be a reduction of performance!
see, for example, IBM AI Fairness 360 for (open
source) demo of different algorithms
Zafar, M.B., I. Valera, M. Gomez-Rodriguez, and K.P. Gummadi (2019). Fairness Constraints: A Flexible Approach for Fair Classification, Journal of Machine Learning Research 20 , pp. 1-42
source: Zafar et al. (2019)
3. Fairness-by-design
Example In-Processing](https://image.slidesharecdn.com/20201205fairness-201208094154/85/Ramon-van-den-Akker-Fairness-of-machine-learning-models-an-overview-and-practical-considerations-18-320.jpg)
Ramon van den Akker. Fairness of machine learning models an overview and practical considerations
- 1. Fairness of machine learning models an overview and practical considerations AI & BigData Online Day 2020
- 2. Agenda Disclaimer: The views and opinions expressed in this document are those of the author and do not necessarily reflect the official policy or position of de Volksbank and Tilburg University. Motivating Examples Operationalizing Fairness 1 3 Response: high-level frameworks2
- 3. 1. Motivating examples Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) Recidivism Prediction Instruments (RPI) used by US courts for pretrial decision-making, parole decisions, and sometimes sentencing. On basis of 7,000 cases 2013-2014 in Broward county, Florida ProPublica’s conclusion is that algorithm is biased against blacks.
- 4. Google (2018). Reducing gender bias in Google Translate. Zou, J. and L. Schiebinger (2018). AI can be sexist and racist — it’s time to make it fair, Nature 559, pp. 324-326. 1. Motivating examples
- 5. 1. Motivating examples Ensign, D., S.A. Friedler, S. Neville, C. Scheidegger, and S. Venkatasubramanian (2018). Runaway Feedback Loops in Predictive Policing, Proceedings of Machine Learning Research 81, pp. 160-171.
- 7. 1. Motivating examples 2016 20182016 20182006 2019
- 8. 2. Response: regulation is coming… 2020
- 9. Jobin, Ienca, Vayena (2019): • Mapped and analyzed the current corpus of principles and guidelines on ethical AI • Identified 84 documents containing ethical principles or guidelines for AI, 88% being after 2016. • Convergence? • “In particular, the prevalence of calls for transparency, justice and fairness points to an emerging moral priority”. Transparency & Explainability Justice and fairness Non-maleficence Responsibility Privacy 2. Meta-analysis: Global Landscape of Ethical Frameworks for AI and data science From point-of-view of data scientist: fairness and explainability appear to most difficult to implement.
- 10. 3. Non-discrimination and data science but for data scientist, typical goal is to maximize discriminatory power! Models should not yield unintended discrimination
- 11. protected features A X Y data model featureslabels3. Avoiding unintended discrimination: forbidden variables Supervised Learning Ignoring protected features does not work (in general)!
- 12. 3. Defining (un)fairness, discrimination How to define (un)fairness? Note that is actually not a new topic see, for example, Hutchinson and Mitchell (2019) rapidly developing literature mainly for binary classifiers! Barocas, S., M. Hardt, and A. Narayanan (2020). Fairness and machine learning – limitations and opportunities. Book in progress (https://fairmlbook.org/) Corbett-Davies, S. and S. Goel (2018). The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning, working paper Hutchinson, B. and M. Mitchell (2019). 50 Years of Test (Un)fairness: Lessons for Machine Learning. In: FAT* ’19: Conference on Fairness, Accountability, and Transparency (FAT* ’19), ACM, pp. 49-58. Verma, S. and J. Rubin (2018). Fairness Definitions Explained. In: 2018 IEEE/ACM International Workshop on Software Fairness (FairWare). IEEE, pp. 1–7. statistical parity fairness through unawareness equal opportunity predictive parity calibration conditional statistical parity equalized odds counter- factual fairness no proxy discrimination
- 13. Great variety of open-source fairness dashboarding tools are available AI Fairness 360 3. Bias/fairness dashboarding Saleiro, P., B. Kuester, A. Stevens, A. Anisfeld, L. Hinkson, J. London, and R. Ghani (2018). Aequitas: A Bias and Fairness Audit Toolkit, arXiv preprint arXiv:1811.05577. How to proceed if protected group concerns special personal data? Implementation requires choices for protected groups and fairness metrics!
- 14. 3. Why do we need to make a choice: just consider all fairness metrics We think ProPublica’s report was based on faulty statistics and data analysis, and that the report failed to show that the COMPAS itself is racially biased, let alone that other risk instruments are biased. -- Flores et al. (2016) There’s software used across the country to predict future criminals. And it’s biased against blacks. -- ProPublica (2016) source: Chouldekova (2017) Chouldechova (2017). Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments, Big Data 5(2), pp. 153-163. Flores, A.W., K. Bechtel, and C.T. Lowenkamp (2016). False Positives, False Negatives, and False Analyses: A Rejoinder to “Machine Bias: There’s Software Used Across the Country to Predict Future Criminals. And It’s Biased Against Blacks, Federal Probation 80(2), pp.38-46. Kleinberg, J., S. Mullainathan, and M. Raghavan (2017). Inherent Trade-Offs in the Fair Determination of Risk Scores. Proc. 8th Conf. on Innovations in Theoretical Computer Science (ITCS). source: Chouldekova (2017) In general it is impossible to satisfy multiple fairness definitions at the same time! A choice is therefore required.
- 15. 3. Fairness-by-design “post” processing “pre” processing “in” processing No changes to model itself Alter use of model or insert additional randomness No changes to algorithm Alter input data Tailor-made methods: new algorithms If we detect problems: how to fix them? And how to create models which are `fair-by-design’?
- 16. Google (2018). Reducing gender bias in Google Translate. suppose 𝑷 𝒀 = 𝟏 | 𝑨 = 𝒃 = 0.3 and 𝑷 𝒀 = 𝟏 | 𝑨 = 𝒘 = 0.5. add additional randomness: for A = b and 𝒀 = 𝟎 flip a coin with prob. `heads’ equal to 0.2 and set 𝒀 = 𝟏 if `heads’ yields statistical parity but loss of performance! Examples Post Processing 3. Fairness-by-design
- 17. protected features A X Y data model features transformed features Calmon, F.P., D. Wei, B. Vinzamuri, K.N. Ramamurthy, and K.R. Varshney (2017). Optimized pre-processing for discrimination prevention, NIPS. advantage: existing algorithms can still be used in Step 2 in general there will be a reduction of performance! Step 1 Step 2 Example Pre-Processing 3. Fairness-by-design
- 18. new tailor made algorithms / model common structure: max “performance” subject to “fairness constraints” max [“performance” + 𝛼“fairness”] in general there will be a reduction of performance! see, for example, IBM AI Fairness 360 for (open source) demo of different algorithms Zafar, M.B., I. Valera, M. Gomez-Rodriguez, and K.P. Gummadi (2019). Fairness Constraints: A Flexible Approach for Fair Classification, Journal of Machine Learning Research 20 , pp. 1-42 source: Zafar et al. (2019) 3. Fairness-by-design Example In-Processing
- 19. 3. Round-up and reflection Implementation fairness requires choice of protected groups and choice for fairness metric Law does not seem to offer guidance on the precise metric GDPR can yield problems if definition protected group depends on special personal data. Even worse: in some cases it will concern data that most companies do not (want to) have available There is trade-off between performance and fairness. If not clearly specificied in law: what is the appropriate balance? Gives rise to new governance questions: - Who should be responsible for decisions on fairness? - How to ensure company-wide consistency in decision making?
- 20. About me Ramon van den Akker Risk Modelling and AI Center, de Volksbank Associate Professor, department of Econometrics & OR, Tilburg University ramon.vandenakker@devolksbank.nl, r.vdnakker@uvt.nl www.linkedin.com/in/ramonvdakker/
Editor's Notes
- #16: 21+ definitions of fairness No consensus on which one is the best, depends on the case Most are mathematically incompatible