一种基于结构化学习的排序算法

传统排序算法将排序问题转换成分类或回归问题来求解，这样得到的模型不够精确。对此提出一种新的排序算法，该算法把排序问题看成一个结构化学习过程，即通过训练集来学习一个排序结构。算法首先定义了一个查询级的目标函数，针对算法约束条件太多，难以直接优化，提出使用割平面算法进行求解。对于算法中的“寻找最违约排列”子问题，将其变换成为一个简单的降序排列问题。基于基准数据集的实验表明，相比起传统的排序算法，所提算法更为有效。

Ranking algorithm based on structured learning

For the problem that the model learned from traditional ranking algorithm which converts ranking problem to classification or regression is not accurate, a novel ranking algorithm is proposed.It views the ranking problem as a procedure of structured learning which learns a rank structure from the train set.The algorithm defines a object function of query level, and presents using the cutting plane algorithm to solve the problem that the algorithm has exponential number of constraints. For the sub-problem of finding the most violated constraints,the paper transforms it into a simple sorting in descending order.Experimental results on the benchmark datasets show that the algorithm proposed in this paper is more effective than the traditional ranking algorithm.