Fairness in Automated Decision Systems
- 1. Fairness in Automated Decision Systems Data Protectionin Software Development- WS 21/22 Dr. AmirshayanAhmadian • Sai Sri Praveen Gadiyaram • Sowjanya Chennamaneni 26-01-2022 1
- 2. OUTLINE 1. Motivation 2. Background 3. Challenges 4. Fairness metrics 5. Fairness approaches 6. Fairness Architecture 7. Adversarial Debiasing 8. Case-study 9. Summary 26-01-2022 2
- 4. BACKGROUND 26-01-2022 AutomatedDecision Systems : Decisionstaken by the systems without any human involvement. 4
- 5. . 26-01-2022 Name Education Race Gender Zip code John Bachelor's White Male 56783 Mary Bachelor's White Female 45673 Annie Bachelor's Black Female 15674 Jack Bachelor's Black Male 13456 TrainedML model Predictions Eligible Yes Yes No No BACKGROUND 5 • Protected Attributes: Can be a source of social bias. For example, age, race and gender • Unprotected Attributes: Attributes that are not protected.
- 6. • Fairness ThroughAwareness: An algorithm is fair if it gives similar predictions to similar individuals. It is also known as Individual Fairness. • DemographicParity : The likelihood of a positive outcome should be the same regardless of whether the person is in the protected (e.g., female) group. It comes under Group Fairness 26-01-2022 6 BACKGROUND
- 7. CHALLENGES • Decisions are dependent on the "Data". • A model trained on biased data can produce unfair and discriminatory decisions. • Algorithm bias 26-01-2022 Interpretabilityof models Universalfairnessdefinition Universalfairnessmetrics Concretebias mitigation Lack of social fairness Data Algorithm Decision 7
- 8. FAIRNESS METRICS 26-01-2022 S. No Metric Definition formula 1. Disparate Impact Ratio(DIR) Ratio of the rate of positive outcome for minority group to majority group P(y=1|x=0)/P(y=1|x=1) 2. Statistical Parity Difference(SPD) Difference of the rate of positive outcome for minority group to majority group P(y=1|x=0) - P(y=1|x=1) • Quantifying Group Fairness • 2 Groups: • Male, x=1 • Female, x=0 8
- 9. FAIRNESS METRICS • Effectively search the attribute space • Test-intensive – Can't be done manually • Generate automatic Test cases for indiscrimination. 26-01-2022 Protected Non- protected Individual discrimination • path coverage and individual discrimination • Always negate a non-protected attribute • Success score= |setof generated testcasesfrom unprotected attributes| |individual discrimination detected| 9
- 10. FAIRNESS APPROACHES • Achieving Fairness == Mitigating Bias • Consider fairness constraints in decision system building • 3 different approaches to include fairness constraints 26-01-2022 Pre-processing • Learn fair Representations • Remove disparate impact In-processing • Regularization • Train independently and tune together • Adversarial debiasing Post-processing • Calibrated Equal Odds 10
- 12. ADVERSARIAL DE-BIASING • Optimize with respect to the fairness loss • Adversarial NN can predict the attributes which leads to an unfair decision. 26-01-2022 Gender ?? Race ?? 12
- 13. CASE-STUDY • Fairness architecture at LinkedIn • "Equal opportunities to equally qualified members" • Model agnostic and scalable • Fair model analyzer: • Specialized test set • Creates the fairness correction • Mitigation trainer • Learns from the previous step output • post-processing approach 26-01-2022 13
- 14. SUMMARY 26-01-2022 • 100 % data-driven systems are not ideal • Focus on Bias • Data source & variety • Regularization • Lack of Fairness knowledge • Social Fairness, ethics • Fairness constraints • Context & diversity • Continuous fairness evaluation • Testing after deployment • Optimize fairness metrics 14
- 15. REFERENCES • Mehrabi, Ninareh & Morstatter, Fred & Saxena, Nripsuta & Lerman, Kristina & Galstyan, Aram. (2019). A Survey on Bias and Fairness in Machine Learning. • Kulshrestha, Ankit and Ilya Safro. “CONFAIR: Configurable and Interpretable Algorithmic Fairness.” (2021). • https://engineering.linkedin.com/blog/2022/a-closer-look-at-how-linkedin-integrates-fairness-into-its-ai-pr • https://www.slideshare.net/godatadriven/fairness-in-ai-ddsw-2019 • https://blog.insightdatascience.com/tackling-discrimination-in-machine-learning-5c95fde95e95 • https://www.persistent.com/blogs/fairness-in-ai-systems/ • Aniya Aggarwal, Pranay Lohia, Seema Nagar, Kuntal Dey, and Diptikalyan Saha. 2019. Black box fairness testing of machine learning models. In <i>Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering</i> (<i>ESEC/FSE 2019</i>). Association for Computing Machinery, New York, NY, USA, 625–635. DOI:https://doi.org/10.1145/3338906.3338937 • Ignore icons created by Parzival’ 1997 - Flaticon • Ethics icons created by Freepik - Flaticon • Testing icons created by surang – Flaticon • https://www.datanami.com/2018/09/05/how-to-build-a-better-machine-learning-pipeline/ • https://www.oreilly.com/library/view/ai-fairness/9781492077664/ch01.html • https://fortune.com/2018/10/10/amazon-ai-recruitment-bias-women-sexist 26-01-2022 15