一种基于潜变量的Ranking模型构造算法

A Ranking Model Constructing Algorithm Based on Latent Variables

现有的Ranking算法获得的模型全部来自训练数据,因为很多模型的有用信息并不能完全从训练数据中得到,因此这样得到的模型不够精确,对此,提出一种基于潜变量的Ranking算法。该算法以结构化SVM为学习工具,将除训练数据外的其他有用信息以潜变量形式引入算法的框架中,并在此基础上定义了面向NDCG的目标函数。针对该目标函数非凸非平滑,首先使用"凹-凸过程"进行逼近,然后用"近似Bundle法"展开优化计算。基准数据集上的实验结果表明:相比完全依靠训练数据的Ranking算法,本文算法获得的模型更为精确。

It is known that the models obtained via the existing ranking algorithms completely come from training data. Because much useful information on the models cannot be gotten from training data, the models are not usually accurate enough. Aiming at the above shortcoming, this paper proposes a ranking algorithm based on latent variables. Firstly, the algorithm uses structural support vector machine as learning tool, and introduces other useful information, except for training data, into algorithm framework as latent variables. On the basis, it defines an object function orienting NDCG. Because the object function is non-convex and non-smooth, this paper utilizes the concave-convex procedure to approximation, and makes use of proximal bundle method to optimizing computing. Experimental results on the benchmark datasets show that the obtained model via the proposed algorithm is more precise than those only via training data.