A listwise approach for learning to rank based on query normalization network

Learning to rank is one of the hotspots in the intersection between information retrieval and machine learning. In the traditional listwise approach for learning to rank based on the neural network, the model predicts the score of each document independently, which cannot reflect the link between those documents associated with the same query. To solve the problem, this paper proposes a new ranking neural network model called Query Normalization Network (QNN). In QNN, normalization is added as a part of the original neural network model to perform the normalization operation for each query sample collection. Through this operation, the prediction scores of documents returned by the same query are also associated with each other. Then, this paper proposes a listwise approach called Optimizing Normalized Discounted Cumulative Gain (NDCG) Query Normalization Network (OptNDCGQNN) which based on QNN and directly optimize the evaluation measure NDCG. OptNDCGQNN use QNN as model and Stochastic Gradient Descent (SGD) as optimization algorithm to optimize an upper bound function of the original loss function, which directly defined according to the evaluation measure NDCG. Experimental results show that OptNDCGQNN has better ranking performance than other traditional ranking algorithms. It also show that when the amount of training data is large enough, OptNDCGQNN can enhance the ranking performance by training deep neural network.