Algorithmic Fairness:From social good to mathematical framework

Views:
 
     
 

Presentation Description

Keynote at ICWSM 2016

Comments

Presentation Transcript

Algorithmic Fairness: From social good to mathematical framework:

Algorithmic Fairness: From social good to mathematical framework Suresh Venkatasubramanian University of Utah

A truly interdisciplinary area:

A truly interdisciplinary area Sociology Political science Law Computer science Economics

Decision by algorithm: Good?:

Decision by algorithm: Good? Dealing with scale Dealing with complexity Dealing with objectivity

Decision by algorithm: Bad?:

Decision by algorithm: Bad? Learned models are complex They are not accountable They might lead to unfair outcomes

Algorithms can’t be unfair!:

Algorithms can’t be unfair! “Data is truth” “Math is not racist” See more at algorithmicfairness.wordpress.com “This is not a CS problem”

How can computer scientists study algorithmic fairness? :

How can computer scientists study algorithmic fairness? Interpretability Accountability Verificatio n The walled-garden view

How can computer scientists study algorithmic fairness? :

How can computer scientists study algorithmic fairness? Can we do fairness-aware machine learning? The integrated view

A mathematical perspective on fairness:

A mathematical perspective on fairness Bake it into how we design algorithms Use it to validate existing algorithms Guide for best practices

A simple problem: classification:

A simple problem: classification ✔ ✖ Hiring College admission Loan

Definitions of fairness:

Definitions of fairness I treat you differently because of your race Individuals with similar abilities should be treated the same Individual fairness Structural bias against groups Groups should all be treated similarly Group fairness

Definitions of fairness:

Definitions of fairness Individual fairness Group fairness

Injecting fairness into algorithmic decision-making:

Injecting fairness into algorithmic decision-making ✔ ✖ Modify the input Modify the algorithm Modify the output [ Romei and Ruggieri 2011?]

Information and computation:

Information and computation

Fano’s inequality:

Fano’s inequality X X’ Y X = X’? Information content Reconstruction error

A key insight: computation to check information flow:

A key insight: computation to check information flow Take data set D containing X Strip out X in some way, to get Y See if we can predict X’ = X from Y with the best possible method. If error is high, then X and Y have very little shared information.

Consequence: a tester:

Consequence: a tester

Disparate Impact:

Disparate Impact “4/5 rule”: There is a potential for disparate impact if the ratio of class-conditioned success probabilities is at most 4/5 Focus on outcome, rather than intent.

Certification via prediction:

Certification via prediction X ? Y Theorem: If we can predict X from Y with probability ε , then our classifier has potential disparate impact with level g( ε ).

Consequences:

Consequences We don’t need to know how the classifier works. We don’t even need to know how Y is generated from X. This argument is independent of the computation. If tools for finding proxies for X get better, tools for detecting that something is a proxy for X also get better at the same time! Need to use balanced error rate: error averaged over classes.

Consequence: repair:

Consequence: repair

Different distributions:

Different distributions

Moving them together:

Moving them together

Using the earthmover distance:

Using the earthmover distance Let We find a new distribution that is “close” to all conditional d istributions.

Moving them together:

Moving them together

Example results:

Example results

Consequence: Audits:

Consequence: Audits

Training vs Testing:

Training vs Testing ✔ ✖

Training vs Testing:

Training vs Testing ✔ ✖ Training data Test data No access to training data or algorithm

Our solution:

Our solution Treat each attribute as the one we want to “repair” with respect to Run model on modified test data Associate resulting accuracy with each attribute, get rank ordered list.

Results:

Results Dark reactions project: predict presence/absence of a certain compound in a complex reaction. 273 distinct features. Approach identified key variables for further study that appear to influence the models.

Challenges:

Challenges

Definitions of fairness:

Definitions of fairness Theories of justice: Rawls and the “veil of ignorance” Nozick and the minarchist state. O thers...? Individual vs group fairness What are the underlying assumptions about the world? H ow do these affect the choice of algorithm?

The ground-truth problem:

The ground-truth problem Mask all information: perfectly fair but useless! Utility based on potentially biased training data! Fairness Utility Less biased data G 1 G 2 G 3 G 4

Fairness and privacy:

Fairness and privacy ✔ ✖ ? ? Hide bias-causing information: GOOD If you can’t see it, you can’t act on it Hide bias-causing information: BAD If it can’t be seen, malicious intent cannot be detected.

Agent interactions:

Agent interactions Is fairness in decision-making an adversarial problem or a cooperative one? How do the mechanisms change? Disparate impact law is adversarial Audits (financial etc ) are not. Intentional discrimination claims are adversarial Hiring practices can be certified as compliant.

Interaction with law/policy:

Interaction with law /policy Discrimination case law all over the place: Hiring Admissions Loans (car, housing) Affirmative action (may be) OK, quota systems are not. “case law not designed keeping algorithms in mind” Mechanisms Laws Is today the day my research is declared illegal?

In closing:

In closing

My Collaborators Ifeoma Ajunwa (UDC) Sorelle Friedler (Haverford) Carlos Scheidegger (U. Arizona) and many more… :

My Collaborators Ifeoma Ajunwa (UDC) Sorelle Friedler (Haverford) Carlos Scheidegger (U. Arizona) and many more … suresh@cs.utah.edu http://fairness.haverford.edu http://fatml.org

authorStream Live Help