一种基于新型损失函数的 Listwise 排序学习方法-《计算机技术与发展》

文章信息/Info

作者:: 龚安¹ ; 孙辉¹ ; 乔杰 ²; 1. 中国石油大学(华东) 计算机与通信工程学院,山东青岛 266580; 2. 中国石油大学(华东) 石油工程学院,山东青岛 266580

Author(s):: GONG An¹ ; SUN Hui¹; QIAO Jie²; 1. School of Computer &Communication Engineering,China University of Petroleum,Qingdao 266580,China; 2. School of Petroleum Engineering,China University of Petroleum,Qingdao 266580,China

Keywords:: sorting learning; loss function fusion; Listwise; gradient descent

摘要:: 排序学习是指运用机器学习方法,自主地构建排序模型,用来对新的数据进行排序。在所有的排序方法中,List- wise 方法就是其中一类重要的排序学习方法,它的训练样例由文档列表组成,利用神经网络模型和概率模型来构造损失函数。但是由于 Listwise 算法存在时间复杂度高、排序位置信息利用度低等缺点,一直得不到广泛的推广。对此,文中在 SHF-SDCG 框架的基础上提出了一种新的排序学习算法,采用多层神经网络的 ListNet 算法,引入 Pointwise 损失函数和位置加权因子,与 Listwise 损失函数融合构建新的损失函数,并分别使用梯度下降算法和多层神经网络算法训练网络权值, 得到新的排序模型;同时使用效率高的 Top-k 训练方法,降低时间复杂度。最后在数据集 LETOR4. 0 上进行实验,结果表明新算法排序性能明显提高。

Abstract:: Learning to rank refers to the use of machine learning methods and builds ranking model autonomously to rank new data. A- mong all the learning to rank methods,the Listwise method is one of the most important methods and its training dataset is composed of a list of documents. And we make use of the neural network model and probability model to construct the loss function. Besides,because of the disadvantages of Listwise such as high time complexity and low ranking location information,it has not been widely used now. So based on the framework of SHF-SDCG,we propose a new learning to rank algorithm using the ListNet algorithm of multi-layer neural network. The new loss function is constructed by combining the loss function of Pointwise with position weighting factor. Then we use the gradient weighting algorithm and the multi-layer neural network algorithm to train the network weight value. Finally,we get the new learning to rank model. At the same time,we adopt the high-efficiency Top-k training method to reduce the time complexity. The ex- periment on the dataset of LETOR4.0 shows that the performance of the proposed algorithm is obviously improved.