Search for the smallest random forest.
Search for the smallest random forest.
复制标题
DOI：
10.4310/sii.2009.v2.n3.a11
复制
发表时间：
2009-01-01
影响因子：
0.8
通讯作者：
Wang M
中科院分区：
文献类型：
--
作者：
Zhang H;Wang M
随机森林已成为许多科学领域最常用的非参数统计方法之一，特别是在高通量基因组数据的分析中。使用随机森林的一般做法是生成足够多的树，尽管对于多大才足够是主观的。此外，随机森林因其庞大的规模而被视为“黑匣子”。在这项工作中，我们解决了使用随机森林的一个基本问题：随机森林必须有多大？为此，我们提出了一种特定的方法来寻找子森林（例如，在个位数的树中），该子森林可以达到大型随机森林（在数千棵树的数量级）的预测精度。我们通过广泛的模拟研究和乳腺癌预后的真实研究对其进行了测试。结果表明，这样的子森林通常存在，而且大多数都非常小，这表明它们实际上是整个随机森林的“代表”。我们得出的结论是，子森林确实是随机森林的核心。因此，没有必要使用整个森林来满足预测性能。此外，通过将随机森林的大小减小到可管理的大小，随机森林不再是黑匣子。
Random forests have emerged as one of the most commonly used nonparametric statistical methods in many scientific areas, particularly in analysis of high throughput genomic data. A general practice in using random forests is to generate a sufficiently large number of trees, although it is subjective as to how large is sufficient. Furthermore, random forests are viewed as “black-box” because of its sheer size. In this work, we address a fundamental issue in the use of random forests: how large does a random forest have to be? To this end, we propose a specific method to find a sub-forest (e.g., in a single digit number of trees) that can achieve the prediction accuracy of a large random forest (in the order of thousands of trees). We tested it on extensive simulation studies and a real study on prognosis of breast cancer. The results show that such sub-forests usually exist and most of them are very small, suggesting they are actually the “representatives” of the whole random forests. We conclude that the sub-forests are indeed the core of a random forest. Thus it is not necessary to use the whole forest for satisfying prediction performance. Also, by reducing the size of a random forest to a manageable size, the random forest is no longer a black-box.