Random forests
In Machine Learning (volume 45, issue 1, October 2001, pages 5-32) Leo Breiman of the University of California, Berkeley, defined random forests: an ensemble of trees, each grown from an independently drawn random vector, voting for the class. He proved that a forest's generalisation error converges almost surely to a limit as trees are added, bounded it by the strength of the individual trees and the correlation between them, and showed that choosing features at random at each split gives error rates that compare favourably with AdaBoost while being more robust to noise.
Why it matters
An ensemble of trees got a theory of when adding trees stops helping and why, together with estimates of error, strength, correlation and the importance of each variable computed from the training data itself. The Kinect body-part classifier of 2011 is a randomized decision forest, and its paper cites this one.
Breiman places the method after bagging (his own, 1996) and Dietterich's random split selection (1998). The comparison with AdaBoost uses 20 data sets: 13 smaller ones from the UCI repository, 3 larger ones with separate training and test sets, and 4 synthetic ones. The full text read is Breiman's report dated January 2001 on his Berkeley page: its abstract matches the journal's word for word, but the figures of the body are the report's and were not checked against the journal typesetting. The journal date, October 2001, is from Springer's page and the Crossref record. What the record does not claim: specific error rates from the tables (not copied), or that random forests became an industry standard, which no source read says.