5 years ago
Wise Practitioner – Predictive Analytics Interview Series: Jeff Heaton at Reinsurance Group of America

By: Luba Gloukhova, Founding Chair, Deep Learning World

In anticipation of his upcoming conference presentation at Deep Learning World Las Vegas, June 16-20, 2019, we asked Jeff Heaton, VP, Data Scientist at Reinsurance Group of America, a few questions about his work in deep learning. Catch a glimpse of his presentation, How Much Data is Enough for Deep Learning, and see what’s in store at the DLW conference in Las Vegas.

Q: In your work with deep learning, what do you model (i.e., what is the dependent variable, the behavior or outcome your models predict)?

I work on a variety of life insurance-related models. The dependent variable might be the propensity for existing customers to buy another product, the probability of an existing policy holder to lapse a product, or even the probability of an individual having a risky behavior such as smoking.

Q: How does deep learning deliver value at your organization – what is one specific way in which model outputs actively drive decisions or operations?

Feature engineering is a critical part (and time-consuming part) of every modeling project. The ability for deep learning technologies to automatically engineer some features is a very valuable capability. Convolutional neural networks can be used to engineer features from time-series data to predict future customer behavior based on their activities over time.

Q: Can you describe a quantitative result, such as the performance of your model or the ROI of the model deployment initiative?

At one point we were able to achieve a +3.1% accuracy boost by including features that were engineered using deep denoising autoencoders. The ability to perform some automatic feature engineering is one of deep learnings greatest strengths. However, it still cannot replace engineered features created by a data scientist or subject matter expert with advanced domain knowledge.

Q: What surprising discovery or insight have you unearthed in your data?

Models will often show interactions between columns that we did not previously consider. Insurance policies can be terminated very simply by policy holder ceasing payment. Such a termination is called a lapse. Predicting the time that an insurance policy may be lapsed is an important predictive modelling application. Most of these models are not simply black boxes. We are able to see interesting insights into the causality of insurance policy lapse behavior.

Q: What excites you most about the field of deep learning today?

The sheer pace at which deep learning is evolving. It is impossible to keep up with even a fraction of the deep learning research papers being published on arXiv. Just watching Google’s progress from AlphaGo to AlphaZero has been fascinating. For AlphaZero to master chess in four hours, without any prior data, is remarkable. In some ways, Google is taking us from “Big Data” to “Infinite Data” in that AlphaZero had as many chess games to learn from as it has had compute time to generate.

Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Deep Learning World.

Though it is an overused buzz word, “Big Data” has revolutionized machine learning. But not all data are created equal. In my talk, I will show you how to measure the extent to which your data is varied and how to determine if new data being scored are covered by your training data. It is easy to check if your new data are outside of the ranges of your individual training data. However, evaluating training set coverage in a holistic, truly multi-variate level, is difficult.