一种基于listwise的在线学习书目排序检索算法

An online learning sorting algorithm based on listwise for book list retrieval

高效检索是数字图书馆的核心业务之一,其中排序是高效信息检索的核心问题。给定一系列的书目列表,利用排序模型生成目标书目的排序列表。将学习排序算法应用于信息检索领域时,常用方法是通过最小化pairwise损失函数值来优化排序模型。然而,已有结论表明,pairwise损失值最小化不一定能得到listwise算法的最佳排序性能。并且将在线学习排序算法与listwise算法相结合也非常困难。提出了一种基于listwise的在线学习排序算法,旨在保证listwise算法性能优势的前提下,实现在线学习排序算法,从而降低检索复杂度。首先解决将在线学习排序算法与listwise算法相结合的问题;然后通过最小化基于预测列表和真实列表定义的损失函数来优化排序模型;最后提出基于online-listwise算法的自适应学习率。实验结果表明,所提出算法具有较好的检索性能和检索速度。

Efficient retrieval is one of the crucial services of digital library, and sorting is the core issue of efficient information retrieval. Given a list of candidate book titles, the sorting model is used to generate a sorted list of book titles. When learning-based sorting algorithms are applied in the filed of information retrieval, they often optimize the sorting model by minimizing the value of the pairwise loss function. However, existing analysis shows that minimizing the pairwise loss value does not necessarily lead to the optimal sorting performance of the listwise algorithm. It is also very difficult to combine the online learning sorting algorithm with the listwise algorithm. This paper proposes an online learning sorting algorithm based on listwise, which aims to realize the online learning sorting algorithm under the premise of ensuring the performance advantage of listwise algorithm, thereby reducing the retrieval complexity. Firstly, the problem of combining the online learning sorting algorithm with the listwise algorithm is solved. Secondly, the sorting model is optimized by minimizing the loss function based on the predicted list and the real list. Finally, an adaptive learning rate based on the online-listwise algorithm is proposed. Experimental results show that the proposed algorithm has better retrieval performance and speed.