代价敏感的列表排序算法

Cost-Sensitive Listwise Ranking Approach

排序学习是信息检索与机器学习中的研究热点之一.在信息检索中,预测排序列表中顶部排序非常重要.但是,排序学习中一类经典的排序算法——列表排序算法——无法强调预测排序列表中顶部排序.为了解决此问题,将代价敏感学习的思想融入到列表排序算法中,提出代价敏感的列表排序算法框架.该框架是在列表排序算法的损失函数中对文档引入权重,且基于性能评价指标NDCG计算文档的权重.在此基础之上,进一步证明了代价敏感的列表排序算法的损失函数是NDCG损失的上界.为了验证代价敏感的列表排序算法的有效性,在此框架下提出了一种代价敏感的ListMLE排序算法,并对该算法开展序保持与泛化性的理论研究工作,从理论上验证了该算法具有序保持特性.在基准数据集上的实验结果表明,在预测排序列表中顶部排序中,代价敏感的ListMLE比传统排序学习算法能取得更好的性能.

Learning to rank is a popular research area in machine learning and information retrieval （IR）. In IR, the ranking order on the top of the ranked list is very important. However, listwise approach, a kind of classical approach in learning to rank, cannot emphasize the ranking order on the top of the ranked list. To solve the problem, the idea of cost-sensitive learning is brought into the listwise approach, and a framework for cost-sensitive listwise approach is established. The framework imposes weights for the documents in the listwise loss functions. The weights are computed based on evaluation measure. Normalized Discounted Cumulative Gain （NDCG）. It is proven that the losses of cost-sensitive listwise approaches are the upper bound of the NDCG loss. As an example, a cost- sensitive ListMLE method is proposed. Moreover, the theoretical analysis is conducted on the order preservation and generalization of cost-sensitive ListMLE. It is proven that the loss function of cost- sensitive ListMLE is order preserved. Experimental results on the benchmark datasets indicate that the cost-sensitive ListMLE achieves higher ranking performance than the baselines on the top of the ranked list.