基于pairwise的改进ranking算法

Improved ranking algorithm based on pairwise method

传统基于pairwise的ranking算法,学习后得到的模型在用NDCG这样的ranking标准评价时效果并不好,对此提出了一种新型ranking算法。该算法也是使用样本对作为训练数据,但定义了一个面向NDCG评估标准的目标函数。针对此目标函数非平滑、难以直接优化的特点,提出使用割平面算法进行学习,不仅解决了上述问题,而且使算法迭代的次数不再依赖于训练样本对数。最后基于基准数据集的实验证明了算法的有效性。

The model learned by ranking algorithm based on traditional pairwise method does not work well by ranking measure,such as Normalized Discounted Cumulative Gain（NDCG）.To solve this problem,a new ranking algorithm was proposed.The algorithm used the same train data as the traditional way,and the difference is defining a new object function directed to NDCG.For the problem that the function is non-smooth,difficult to directly optimize,it was proposed to use the cutting plane algorithm,which not only solved the problem above but also made the number of iteration not depending on the training size.The experimental results on the benchmark datasets prove the effectiveness of the proposed algorithm.