In this special guest feature, Ilya Gerner, Director of Compliance Strategy at GCOM, explains why bias matters when using artificial intelligence (AI) for fraud detection. By understanding the key concepts of machine learning (ML), an organization can increase the fairness of his AI output. Ilya has over 10 years of experience in advanced analytics, leading teams in developing fraud detection algorithms, building decision support tools, and performing statistical analyses. Supporting the Internal Revenue Service’s Identity Theft Strategy Initiative and Security since 2020, he led efforts to provide data analytics capabilities to Summit and the Information Sharing and Analysis Center (ISAC) to identify gaps in identity theft prevention. We conduct a strategic risk analysis for .
Corruption can be a major problem for government agencies that provide benefits to the public. As an example, the rate of unemployment benefits improperly paid by the state can exceed 40%.
Artificial intelligence (AI) can help. AI can comb through large amounts of data to uncover potential fraud, and it can do so much faster and more accurately than humans. So it’s no wonder more and more agencies are turning to AI to identify fraudulent activity and reduce fraudulent payments.
However, AI is known to be subject to bias. For example, the Lensa AI image generator was recently found to provide renderings that alter the appearance of people in ways that are considered biased based on their gender and race.
Bias can enter a machine learning (ML) model in a number of ways. One way is to use historical data. When you train a model on a dataset that itself has a bias, that bias is built into the model.
Another method is the introduction of proxy data. For example, imagine you’re looking for evidence of fraud on your tax returns. Even a model that omits the age of a return filer but includes the total number of tax returns filed by that person before may still be a rough approximation of the number of tax returns filed in a lifetime. Therefore, age-based effects may still differ. by age.
Inequity is a particular concern for governments working with datasets containing legally protected attributes such as age, gender and race. Agencies want to avoid both disparate treatments that apply decisions to demographic groups in different ways and disparate effects that harm or benefit demographic groups in different ways.
However, strategies exist to avoid bias and unfairness in AI fraud detection. By gaining a deeper understanding of how ML models work, organizations can ensure AI fairness.
ML concepts for de-biasing
Four major approaches to mitigating unfairness in ML algorithms (unawareness, demographic parity, equalized odds, and predicted rate parity) and how they are deployed in hypothetical real-world scenarios. Let’s see if
State tax authorities make every effort to collect delinquent taxes, but due to resource constraints, not all cases can be resolved. As a result, agencies prioritize deals that generate large amounts of funds at low cost.
Suppose the state tax department wants to identify 50 taxpayers and mail them a notice that their tax due is overdue. But the company wants to avoid contacting taxpayers, who could consume resources by calling the agency’s customer service center after receiving the notice.
The agency knows from historical data that taxpayers over the age of 45 are more likely to call customer service centers. So age, a sensitive attribute, is also part of the problem. This has different implications depending on which strategies are applied to mitigate injustice.
Mitigation Approach 1: Lack of Awareness. Models using this concept omit sensitive attributes such as age. However, delegation of such sensitive attributes is not considered.
In this hypothetical example, the tax authority’s ML model applies the concept of unawareness to select taxpayers based on the frequency of calls to the contact center. Although we don’t directly use age as an attribute, age correlates with phone usage, which favors younger taxpayers. As a result, the model selects 35 taxpayers under 45 and 15 taxpayers over 45. As a result, 10 taxpayers will call the contact center. This is not a bad result, but probably not ideal.
Mitigation Approach 2: Population Equity. Using this concept, the model’s probability of predicting a particular outcome is the same for one individual as for another individual with different sensitive attributes.
Applying the concept of population equality, the tax authority’s ML model uses age directly to ensure an equal distribution of taxpayers over 45 and under 45. As a result, the model selects 25 taxpayers under 45 and 25 taxpayers over 45. As a result, he will be 14 people. Taxpayers call the contact center, but the results are less favorable than the concept of ignorance.
Mitigation Approach 3: Even out the odds. With Equalized Odds, if the model predicts the same outcome for two individuals with different sensitive attributes, then one of them has the same probability of being selected.
Applying this concept, the agency’s ML model uses age to ensure that the true positive and false positive rates are the same for taxpayers over 45 and under. As a result, 30 taxpayers under 45 and her 20 taxpayers over 45 are selected. In this case, eight taxpayers called the customer service center, the best result so far.
Mitigation Approach 4: Predicted Rate Parity. The concept is that if a model predicts a particular outcome for two individuals with different sensitive attributes, both individuals have the same probability of predicting that outcome.
Applying this concept, the agency’s ML model uses age to ensure that the number of taxpayers calling the contact center is under 45 and over 45. As a result, 40 taxpayers under the age of 45 and 10 taxpayers over the age of 45 are selected. Call the Customer Service Center – same result as for Equalized Odds.
Summarizing the results for this hypothetical situation, two of the models achieve better results, but rely on sensitive data. A model that doesn’t rely on sensitive data achieves fairly desirable results, but the data used is a proxy for sensitive data.
balance accuracy and fairness
One of the challenges for ML modelers is that the four concepts for reducing unfairness are mutually exclusive. A modeler must choose one fairness definition to apply to the algorithm and accept the tradeoffs.
Demographic parity, equalized odds, and predicted rate parity all have different treatments. If you are unaware, you will not be treated differently, but it can have a different impact. Each concept has strengths and weaknesses, and there is no right or wrong choice.
Another challenge is that there is often a trade-off between accuracy and fairness. A very accurate model may not be fair. However, improving model fairness can reduce model accuracy. For fraud detection, an agency may choose to run a less accurate model in order to make the fraud detection fairer.
AI can help governments identify and prevent fraud more efficiently and effectively. Importantly, we understand how ML concepts can impact treatment and outcomes, and we are transparent about how AI is used. Employing strategies to avoid bias and unfairness in AI-powered fraud detection can serve the public fairly.
Sign up for the free insideBIGDATA newsletter.
Join us on Twitter: https://twitter.com/InsideBigData1
Join us on LinkedIn: https://www.linkedin.com/company/insidebigdata/
Join us on Facebook: https://www.facebook.com/insideBIGDATANOW
