Plackett-Luce gradient estimation enables the optimization of stochastic ranking models within feasible time constraints through sampling techniques. Unfortunately, the computational complexity of existing methods does not scale well with the length of the rankings, i.e. the ranking cutoff, nor with the item collection size. In this paper, we introduce the novel PL-Rank-3 algorithm that performs unbiased gradient estimation with a computational complexity comparable to the best sorting algorithms. As a result, our novel learning-to-rank method is applicable in any scenario where standard sorting is feasible in reasonable time. Our experimental results indicate large gains in the time required for optimization, without any loss in performance. For the field, our contribution could potentially allow state-of-the-art learning-to-rank methods to be applied to much larger scales than previously feasible.
翻译:Plackett-Luce 梯度估计有助于在可行的时间范围内,通过抽样技术优化随机排序模型。 不幸的是,现有方法的计算复杂性与排名长度(即分级临界值或项目收集规模)不相称。 在本文中,我们引入了新型的PL-Rank-3算法,该算法使用与最佳排序算法相类似的计算复杂性来进行公正的梯度估计。 因此,我们的新颖的逐级排序方法适用于任何在合理时间内标准排序是可行的情况。 我们的实验结果表明,在优化所需的时间里取得了巨大进步,而没有造成任何绩效损失。 对于外地来说,我们的贡献有可能允许将最先进的学习排序方法应用到比以往可行的大得多的规模上。