Web-Search Ranking with Initialized Gradient Boosted Regression Trees

In May 2010 Yahoo! Inc. hosted the Learning to Rank Challenge. This paper summarizes the approach by the highly placed team Washington University in St. Louis. We investigate Random Forests (RF) as a low-cost alternative algorithm to Gradient Boosted Regression Trees (GBRT) (the de facto standard of web-search ranking). We demonstrate that it yields surprisingly accurate ranking results — comparable to or better than GBRT. We combine the two algorithms by first learning a ranking function with RF and using it as initialization for GBRT. We refer to this setting as iGBRT. Following a recent discussion by Li et al. (2007), we show that the results of iGBRT can be improved upon even further when the web-search ranking task is cast as classification instead of regression. We provide an upper bound of the Expected Reciprocal Rank (Chapelle et al., 2009) in terms of classification error and demonstrate that iGBRT outperforms GBRT and RF on the Microsoft Learning to Rank and Yahoo Ranking Competition data sets with surprising consistency.