Thought Leaders in Artificial Intelligence: Stuart Nisbet, Chief Data Scientist, Cadient Talent (Part 2)

Sramana Mitra: It sounds like there are a few areas where you’re using AI and machine learning. If you could start isolating them and discuss one by one, we can dialogue on each of them. 

Stuart Nisbet: In the area of hourly hiring, applicants come to a website and they apply for a job or they may walk in. There are several sources for where an applicant may come. There are employee referrals as well.

We have data for hundreds of millions of records of applicants over the last twenty years and from those applications, we have information about who was hired. We have a system for monitoring a minority of the applicants once they are with the company.

Let’s assume that we have 20 applicants for a job and we choose one of them. What were the inputs for that decision? Did that person last 30 to 180 days or more? If you use their tenure as a basis for how good a hire it was, you assume that you keep good employees and you assume that employees that leave after 15 days were perhaps not as good a hire.

Turnover is accepted especially in the area of hourly hiring, but it certainly is beneficial to try to minimize that. We apply machine learning algorithms. Most people in the field are familiar with a lot of the decision trees, neural networks, and support vector machines.

I don’t want to make it too technical, but we apply eight or nine different methods to determine a champion model. What that model tries to do is to look at past hiring that resulted in long tenure employees and build a model against which we can test with new employees or a 30% holdout from the test environment.

Once we’ve trained those models and we can pick champion models, then we can test them and say with a degree of certainty that we can improve the choices that are made by managers. 

Sramana Mitra: Are you calling the champion model a successful hire?

Stuart Nisbet: No, the champion model would be, in some cases, a decision tree or algorithm that would yield the highest accuracy in terms of predicting a long-tenured employee. In some cases, it could be a gradient-boosting algorithm.

The model that it is built uses a variety of algorithms. We are not just using one single algorithm to determine who is going to be a long-tenured employee. Let’s say you have a 90% accuracy with one model, then for this month, gradient boosting or decision tree is the champion model so that’s the one that we would use for hiring in the next week.

It’s an ever-evolving system. Sometimes a combination of an ensemble of models would be used to pick the candidates. The champion model is simply the model that yielded the best results. That’s not static; it’s always changing based on new hires and economic environments.

It would be interesting for me as a data scientist to see what impact the coronavirus will have on hiring. For a lot of our clients, the number of applicants that they have gotten has skyrocketed. There are a lot of people applying for certain types of jobs and a lot of companies that are thriving in this environment just because their services are dearly needed.

Other areas on the other hand, are suffering. They are having a hard time – especially those businesses in the service industry. It will be interesting to see because the models will change over time. The champion model is just the one that yields the best prediction.

If you think about linear regression, logistic, and different types of models, the champion model is the one that yields the best predictor of your next best employee.

Sramana Mitra: What is the nature of the dataset that you are applying your algorithm to?

Stuart Nisbet: At its very core, the data that we use is an operational system that takes in information. I want to be very clear in terms of information. Personally identifiable information such as name, address, ethnicity, or gender come in through an application or an interview process.

The data is the applicant’s information. The information used to train the models to build the models is information about experience, educational achievement, cognitive assessment, and personality assessment.

Those are geared towards determining whether an applicant is applying for a cashier position, for example, that has a lot of customer interaction vs someone applying as a stock clerk or a nighttime security guard position where there would be far less customer interaction.

Certain personality traits make one person more ideally suited for a job than the other. There are assessments done on that. The nature of data is the information that is provided by an applicant either through their application, assessment, background check, or a study of their work or educational history. This information provides a profile.

When we talk a little bit later about the differentiators, I want to focus on some of the personally identifiable information that we’re able to fully exclude from any of the models and transparently demonstrate how that is done as a benefit to an in person interview. The algorithm has no bias unless we provide that bias to it.

This segment is part 2 in the series : Thought Leaders in Artificial Intelligence: Stuart Nisbet, Chief Data Scientist, Cadient Talent
1 2 3 4 5 6