book

AI Fairness

by Trisha Mahoney, Kush R. Varshney, Michael Hind

April 2020

Intermediate to advanced

34 pages

43m

English

O'Reilly Media, Inc.

Read now

Unlock full access

Are Human Decisions Less Biased Than Automated Ones?AI Fairness Is Becoming Increasingly CriticalDefining FairnessWhere Does Bias Come From?Bias and Machine LearningCan’t I Just Remove Protected Attributes?Conclusion
Tools and TerminologyTerminologyWhich Metrics Should You Use?Individual Versus Group Fairness MetricsWorldviews and MetricsDataset ClassTransparency in Bias MetricsExplainer ClassAI FactSheets
Most Bias Starts with Your DataPre-Processing AlgorithmsIn-Processing AlgorithmsPost-Processing AlgorithmsContinuous Pipeline Measurement
Step 1: Import StatementsStep 2: Load Dataset, Specify Protected Attribute, and Split Dataset into Train and TestStep 3: Compute Fairness Metric on Original Training DatasetStep 4: Mitigate Bias by Transforming the Original Dataset Step 5: Compute Fairness Metric on Transformed Dataset
The Future of Fairness in AI

Content preview from AI Fairness

Chapter 4. Conclusion

The Future of Fairness in AI

Lack of trust in machine learning may be the single greatest factor that prevents AI from reaching its full potential. To build this trust, data scientists must consider measures beyond predictive accuracy. Mitigating bias and ensuring fairness, along with providing robustness and transparency, are essential to ushering in a world in which machine learning systems are partners that humanity can count on.

Fairness is difficult to define for a given application domain and is not something that data scientists should be tackling alone. The inclusion of multiple stakeholders and perspectives is essential before any modeling is undertaken. After some level of consensus is reached on the desirable values and goals for an AI system, the tools provided in AIF 360 can be utilized to detect, understand, and mitigate unwanted bias. These tools address only a narrow sliver of overall fairness issues, and they should be used in conjunction with procedural and organizational measures.