Responsible Artificial Intelligence (RAI) Intro and an Example Issue: Outliers

The goal of worthy AI and data science endeavors is to discover new true things. Every stage of an analytics challenge is susceptible to error and misdirection, seeping in to weaken or destroy useful results.

It takes expertise and discipline—responsible AI (RAI) practices—to guard against these hazards. RAI best practices ensure that analytic decision systems are designed expertly and ethically for greater effectiveness, reliability, and acceptance. Elder Research has survived and thrived its ~30 years by figuring out RAI and delivering high-value projects driven by it.

Everyone at Elder Research, whether technical or not, completes a short course based on our RAI framework, as we collaborate to deliver world-class results. The course doesn’t teach anyone how to do all the essential RAI tasks; that takes many years to master. Instead, for each stage of an analytics project, it thoroughly summarizes the major error entry points that one must defend against.

One such key hazard is outlier data points, which often harm model building; they can reverse, obscure, or even be the important result. I’ll illustrate each of those outcomes through real-world examples:

The Social Security Administration handles disability payments, deciding which applicants qualify as both poor and sick enough to receive taxpayer funds. The latter judgment (degree of disability) is the most challenging for adjudicators. Elder Research was tasked with building models to score each application to considerably improve the speed, accuracy, and consistency of judgements (leaving the hardest cases to human experts).

One important subset of cases had to do with premature infants. Early data exploration found that weight at birth was positively correlated with the target variable (needing assistance: 1: yes; 0: no). This is obviously wrong; it’s the tiniest babies whose families need the most help!

The cause of this misleading finding? An 800-lb baby¹ coded as 1. Remove that outlying case and the model for the other 20,000 cases has the reverse relationship. The problem with the outlier’s impact can be obvious to someone familiar with the human context but not yet to an automated tool working without human intervention.²

Retailers seek useful patterns from point-of-sale data merged with loyalty card data. But the best customers can lead models astray. Huge purchasers are rarely single customers; they may be professional shoppers, institutions, or overly helpful clerks lending their badges to any shopper lacking one.

The sales of the outliers are valid, and they will contribute well to some results (e.g., restocking predictions) but not to others (such as average customer spend by season). Make sure the cases employed are appropriate for the model goal.

Some outliers are discoveries that prove more important than the intended original model. I learned a behind-the-scenes example from Scottish statistician John Aitchison³, and its implications continue today.⁴

British researchers at an Antarctic station noticed a few remote instruments for measuring sun radiation were maxed out. They assumed the outliers were a hardware error and replaced the equipment. They did this again when the devices re-saturated, but the third occasion led them to broaden their hypotheses.

The researchers eventually discovered a “hole” in the ozone layer of our stratosphere, which is dangerous (skin cancer). Their work led to the banning of chlorofluorocarbons (CFCs), and eventually to receipt of a Nobel prize for those who first hypothesized the danger.

The above examples reveal ways a single rogue case in a dataset can have enormous influence over its model. And that issue is just one of hundreds of addressed by RAI!

In each scenario, human creativity and context knowledge were essential to discovering the truth. I believe such expertise and creativity will always be needed to get the best answers, even considering AI’s impressive and growing capabilities.

I will write soon about other key data questions addressed by RAI, including 1) What’s missing from your database? and 2) Are some cases mislabeled?

Footnotes

¹ Was this due to a missing decimal point, or perhaps measuring in grams instead of pounds? (I know a baby who survived at ~1 lb. or ~500 grams.)

² AI can easily highlight possible outliers according to the data (extremes on distributions, for example), which would work well in this example. But I don’t know of a tool that can currently reason about model implications using outside information.

³ https://en.wikipedia.org/wiki/John_Aitchison

⁴ Prof. Aitchison recounted this story the year he started the Statistics Department at the University of Virginia. I had the great fortune of learning much from him. I only learned recently about the Nobel.