Stratified Gradient Boosting for Fast Training of Conditional Random Fields ⋆
Bernd Gutmann, Kristian Kersting · Lirias · 2007
Abstract. Boosting has recently been shown to be a promising approach for training conditional random fields (CRFs) as it allows to efficiently induce conjunctive (even relational) features. The potentials are represented as weighted sums of regression trees that are induced using gradient tree boosting. Its large scale application such as in relational domains, however, suffers from two drawbacks: induced trees can spoil previous maximizations and the number of generated regression examples can become quite large. In this paper, we propose to tackle the latter problem by injecting randomness into the regression estimation procedure by subsampling regression examples. Experiments on a real-world data set show that this sampling approach is comparable with more sophisticated boosting algorithms in early iterations and, hence, provides an interesting alternative as it is much simpler to implement. 1