课程: Advanced Predictive Modeling: Mastering Ensembles and Metamodeling

今天就学习课程吧！

今天就开通帐号，24,600 门业界名师课程任您挑！

Model search by bumping

Model search by bumping - SPSS教程

课程: Advanced Predictive Modeling: Mastering Ensembles and Metamodeling

开始一个月试用 购买团队方案

Model search by bumping

“

- [Keith] Now let's talk about an algorithm which isn't very famous, but I think deserves more attention. Bumping is actually an acronym, is stands for bootstrap umbrella of model parameters, which, frankly, is a pretty fancy name for what this is. It's very straightforward. It's proposed by Tibshirani and his co-author. And of course, I think you'll recognize Tibshirani's name, who's very famous in this space. So here's the basic observation that he's making. He says, "Unfortunately, the averaging process "that produces the bagged estimates "also destroys any simple structure "that's present in the original. "In other words, the bagged subset "is no longer a subset, "and the bagged tree is not a tree." So here's his idea, and it couldn't be easier. He uses the bootstrap samples to search for the best tree, but then simply uses the tree, not an average of the trees. He just simply uses the best tree. The problem is, is none of this would be very useful unless you had a way to…

内容