论文部分内容阅读
鲁棒性在排序学习中显得越来越重要,而现有排序学习算法多数仅关注改进排序模型的有效性,往往忽略了排序模型的鲁棒性.为了增强排序模型的鲁棒性,在训练排序模型的过程中可同时考虑其有效性和鲁棒性.从一个新颖的视角,即偏差-方差均衡,研究了如何优化LambdaMART排序学习的有效性和鲁棒性均衡.将偏差和方差融合为一个统一的目标函数以修改LambdaMART算法中的梯度,并证明了修改后的梯度仍可采用LambdaMART算法去优化以训练排序模型.最后,在排序学习数据集上的实验结果表明,基于偏差-方差均衡思想所修改梯度后的LambdaMART算法具有更强的鲁棒性.
Robustness is more and more important in rank learning, and most of the existing ranking learning algorithms focus only on improving the efficiency of the ranking model, and often ignore the robustness of the ranking model.In order to enhance the robustness of the ranking model, In the process of sorting the model, its validity and robustness can be considered simultaneously.From a novel perspective, that is, deviation-variance equilibrium, we study how to optimize the validity and robustness of LambdaMART scheduling learning.We divide the deviation and variance into A unified objective function to modify the gradient in the LambdaMART algorithm and prove that the modified gradient can still be optimized using LambdaMART algorithm to train the ranking model.Finally, experimental results on a sorted learning dataset show that based on the deviation-variance equilibrium The modified LambdaMART algorithm has more robustness.