You are hereFeature-Weighted Linear Stacking
Feature-Weighted Linear Stacking
In this paper by Joe Sill, Gabor Takacs, Lester Mackey and David Lin of The Ensemble, a blending technique called feature-weighted linear stacking (FWLS) is presented. FWLS achieves significantly higher accuracy levels than standard linear blending while retaining the virtues of linear regression regarding speed, stability, and interpretability. The algorithm produces a linear combination of models using coefficients which are themselves linear functions of meta-features such as the number of user ratings, the number of movie ratings, and the standard deviation of the user ratings. The paper describes a total of 24 such meta-features. FWLS was an important factor in the ascent of Grand Prize Team on the Netflix Prize leaderboard and also contributed significantly to the final submission of The Ensemble, which tied the submission of the winning team (BellKor's Pragmatic Chaos) in terms of accuracy.
Download the latest revision of Feature-Weighted Linear Stacking in PDF and other formats from arXiv.
arXiv is an open access electronic archive and distribution service operated by the Cornell University Library.