A New Tool for Data Mining Salford Systems' Random Forests technology branches out

摘要

Random forests (RF) is a Monte Carlo process in machine learning that was developed by Leo Brei-man of Berkeley to improve on his classification and regression tree (CART) approach to data mining. In essence, RF generates an ensemble of many CARTs using a bootstrap procedure and bags (bootstrap aggregates) over those many trees. Samples of variables are searched at each node rather than investigating all of the variables. The predictive ability is measured on an out-of-bag (OOB) sample of about 37 percent, which is passed down the branches of the trees to obtain class prediction estimates.