Shift AI 2020: How to identify and treat biases in ML Models | Navdeep Sharma (Accenture)
0 likes133 views
The document discusses algorithmic fairness and the importance of addressing bias in machine learning, emphasizing that data biases must be accounted for when designing algorithms. It outlines various definitions of fairness, relevant metrics, and the challenges faced in practical applications, advocating for a collaborative approach that includes both data scientists and business stakeholders. The Accenture fairness tool is introduced as a means to integrate fairness into the data science workflow, allowing for interaction and analysis of potential biases in decision-making processes.
Shift AI 2020: How to identify and treat biases in ML Models | Navdeep Sharma (Accenture)
- 1. Algorithmic Fairness Navdeep Sharma Data & AI Architect - Accenture, The Dock Shift AI
- 2. Algorithmic Bias: Why is it important? Set the scene: Fairness Definitions Introduction to Fairness Metrics Introduction to Accenture Fairness Tool 0201 03 04Agenda
- 3. Our society's growing reliance on algorithmic decision making, particularly in social and economic areas has raised a concern; that they may inadvertently discriminate against certain groups. “Business needs to consider society as a stakeholder.” - Cennydd Bowles, Future Ethics Introduction | Context Objective decision-making is a challenge. “Algorithmic Fairness is a practice that aims to mitigate unconscious bias against any individual or group of people in Machine Learning.”
- 4. “Data biases are inevitable. We must design algorithms that account for them. " “The model summarizes the data correctly. If the data is biased it is not the algorithm’s fault.” VS From “Tutorial: 21 fairness definitions and their politics” on YouTube From “Tutorial: 21 fairness definitions and their politics” on YouTube
- 6. Definitions: “The Treaty on the Functioning of the European Union prohibits (TFEU) discrimination on grounds of nationality. It also encourages to combat discrimination based on sex, racial or ethnic origin, religion or belief, disability, age or sexual orientation.” Status Definition: Privileged vs Unprivileged Example: In criminal risk assessment tools, a common example of protected feature is race, with the designated levels: white defendants – privileged group vs black defendants – unprivileged. Group Bias VS Individual Bias:Protected feature: “Group fairness approaches partition the population/sample into groups and seeks to equalize a statistical measure across the groups. ” “Individual fairness seeks to understand if similar individuals are treated similarly irrespective of their membership to any of the groups.”
- 7. Mutual Information Identifies proxies for protected feature Prevalence Analysis Fraction of a population that satisfies a given outcome Disparate Impact Quantifies disparity of outcomes for different protected groups Predictive Parity - False Positive Rates Proportion of all negatives that still yield positive test outcomes Predictive Parity - True Positive Rates Predictive Parity - Positive Predictive Power Predictive Parity - FalseNegative Rates 7 Metrics Introduction Myriad of metrics: which one to choose? Individualfairness … “Tutorial: 21 fairness definitions and their politics” on YouTube
- 8. Mutual Information Approach: Quantifies the amount of information obtained about one random variable through observing the other random variable. Mutual Information for a Protected Variable, assesses the relationship between the protected and unprotected variables (which could be used in the model build as proxies for sensitive ones and generate bias). Objectives: • Identify proxies • Provoke further analysis: • Is ‘blindness’ w.r.t. protected feature enough? • Predictive power of the proxies with respect to targeted model
- 9. Prevalence Approach: The prevalence of a certain outcome in a certain population can be defined as the fraction of that population that satisfies a given outcome, say Y = reoffended. In other words Prevalence Ratio for a given protected variable is defined as the ratio of the prevalence in the privileged population to the prevalence in the unprivileged group. Prevalence Ratio is calculated on the ground truth. For example: Prevalence ratio (White vs Black) = !"# !$$% %!&" '(&" = 34%
- 10. Disparate Impact TN= 2812 FP = 189 FN = 677 TP = 681 Black Total Pop = 4359 Recidivated Predicted: low risk Predicted: high risk Approach: Unintentional bias is encoded via disparate impact, which occurs when a selection process has widely different outcomes for different groups, even as it appears to be neutral. Calculation: Disparate Impact for a protected variable is the ratio of the % privileged population with a predicted outcome to the % unprivileged population with a predicted outcome. US Law: Originally, the Uniform Guidelines on Employee Selection Procedures provided a simple "80 percent" rule for determining that a company's selection system was having an "adverse impact" on a minority group. Cautionary points: “Courts in the U.S. have questioned the arbitrary nature of the 80 percent rule” TN = 1886 FP = 26 FN = 147 TP = 94 Didnotrecidivate Recidivated Predicted: low risk Predicted: high risk Didnot recidivate White Total Pop = 2154 DI = 0.06/0.2=30% (189+681)/ 4358 = 0.20 (26+94)/ 2154 = 0.06
- 11. False Positive Rate (FPR) TN= 2812 FP = 189 FN = 677 TP = 681 Black Total Pop = 4,359 Recidivated Predicted: low risk Predicted: high risk Approach: Parity for False Positive Rates (FPR) implies that the false positive rates are equal among the privileged and unprivileged population. Calculation: Error Ratio for a given protected variable, is defined as the ratio of error in the privileged population to the ratio of error in the unprivileged population Legislation: No legal precedent for error ratio, however a similar approach to DI can be applied by using the 80% rule (bias when ratio of rates <= 0.8) Didnotrecidivate Recidivated Predicted: low risk Predicted: high risk Didnotrecidivate White Total Pop = 2,154 TN= 1886 FP = 26 FN = 147 TP = 94 189/(2812 + 189) = 0.063 26/(2154 + 26) = 0.013 Error Ratio = 0.013/0.063 = 0.21
- 12. There is now a consensus amongst academia and scientists that algorithmic fairness can not be achieved by the application of data science alone. The complexity of choosing the right solution to allow for group and for individual fairness for example, or to account for accuracy versus fairness, or the complexities that come with scale, and many more is a challenge in itself. All of this is further compounded by myriad of non science factors: What may be fair statistically quite often can fall short ethically or may not be viable from a business perspective. "Bias is a feature of statistical models. Fairness is a feature of human value judgments.” Learnings https://www.semanticscholar.org/paper/Fairness-aware-machine-learning%3A-a-perspective- Žliobaitė/69c7bf934e9ac7673be590f7656bcb38fcb9da48
- 13. What are main general challenges we have encountered when assessing real-life use cases for potential bias? • Metric selection • Academic – Industry gap • Non-binary Protected feature • More than one protected feature • Legislation and guidelines
- 14. Data scientists can solve for many fairness problems from a technical perspective by using statistical metrics but this is not just a data science problem, it requires input from the broader organisation. The tool starts with the data scientist and is integrated with Jupyter hub. We want to add fairness as a step into the current data science workflow. Analyses are pushed to a repository for business users. The business user can explore the interactive analyses and embed them in reports for dissemination for the broader business for decision making. As a communication tool it facilitates a deeper understanding of the challenge. How does the tool work? Accenture Fairness Tool