Learning To Rank Algorithm Based On Deep Belief Network

With the rapid development of the Internet,information retrieval system has become an indispensable tool in people's lives.As the core technology of information retrieval system,ranking learning algorithm has become a research hotspot in the field of information retrieval.Learning to rank is an information retrieval model based on machine learning.Many kinds of machine learning algorithm have been applied to learning to rank,including algorithms based on support vector machine,decision tree and neural network.The ranking algorithm based on neural network have some disadvantages such as the insufficient learning ability and common ranking performance due to the limit of layers.Therefore,we present a new learning to rank algorithm based on the deep belief network,which is called Deep Belief Rank Network(DBRank).The Deep Belief Rank Network algorithm consists of three steps.Firstly,we construct the structure of DBRank with stacking restricted Boltzmann machine(RBM),and use nonsupervised learning method to train it layer by layer in greedy way.Secondly,we construct a pairwise layer as the output layer upon the RBM network,and then use supervised learning method to train it based on pairwise loss function.Thirdly,in order to deal with the shortages including overfitting and disappearance of gradient in the application of neural network,we regularize the back propagation by truncated graduent.Finally,we evaluate the performance of our algorithm from various aspects.we compare our approach with other traditional methods,which shows the outperformance of our method.Besides,the comparison between various numbers of layers and various phases of DBRank shows the validation of pre-training and fine-tuning.