The Journal of Machine Learning Research
| Hi-index | 0.00 |
It has recently been shown how to extend successfully decision tree induction algorithms to bipartite ranking [1]. The major drawbacks of tree-based prediction rules, instability and lack of smoothness namely, are however exacerbated by the global nature of the ranking problem. It is the purpose of this paper to show how to adapt the “bagging” approach, originally introduced in the classification/regression context [2], in order to improve the performance of tree-based ranking rules with regard to these disadvantages. Whereas the notion of majority voting scheme applies to a local prediction problem such as classification or regression in a natural fashion, it is much less straightforward to determine how to average the orderings predicted by many ranking trees. Here we propose various strategies for bagging tree ranking rules inspired by recent advances in the field of rank aggregation for the Web. Strong empirical evidence supporting the fact that they may drastically reduce the variability of unstable statistical procedures such as the TREERANK method is also provided through a simulation study.