Cost-Sensitive Listwise Ranking Approach

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.