Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
This lecture explores a novel approach to stabilizing model selection in data analysis. Learn how to address the instability problem in model selection, where removing a single data point can lead to completely different models being chosen. Rebecca Willett presents a new methodology combining bagging with an "inflated" argmax operation that selects multiple well-fitting models with theoretical stability guarantees. The approach ensures that even when training data points are removed, the resulting model collections will likely overlap with the original selection. See this method demonstrated through real-world applications including ecosystem competition modeling and graph estimation using proteomics cell-signaling data, where it delivers stable, compact, and accurate model collections that outperform existing benchmarks. The presentation covers joint research work with Melissa Adrian and Jake Soloff.
