Smart Subsampling for Predictive Modeling: A Decade of Progress in Material Science (2026-EU-30MP-2798)

Even in the age of powerful machine learning, handling large data sets still comes with its share of challenges. Going brute force – processing everything just because we can – sometimes leads to inefficiencies, overfitting, noise, and rapidly diminishing returns. Smarter doesn’t always mean bigger. Intelligent subsampling, where only a representative fraction of data is used, often reveals clearer patterns and delivers deeper insights.

This topic has fascinated me for nearly 10 years and has been shared several times at this conference. Yet this year marks a major step forward: brand-new results obtained from a much larger data set – roughly 30,000 observations – confirm and extend the power of a smart subsampling strategy. From material science to biomedicine, environmental modeling, marketing, and social sciences, the message remains the same: why go brute force when you can go smart?

Leveraging JMP Pro’s latest capabilities, this work illustrates how combining intelligent subsampling with classical and modern modeling approaches (MLR, Boosted Tree, Neural Nets, SVM, GLM, etc.) leads to robust predictions – and, hopefully, an enjoyable demonstration for all.