Learning to rank with deep neural networks

Details

Supervisors

Dupont, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Due to the growing amount of available information, learning to rank has become an important research topic in machine learning. In this thesis, we address the issue of learning to rank in the document retrieval area. Many algorithms have been devised to tackle this problem. However, none of the existing algorithms make use of deep neural networks. Previous research indicates that deep learning makes significant improvements on a wide variety of applications. Deep architectures are able to discover abstractions, with features from higher levels formed by the composition of lower level features. The thesis starts with the analysis of shallow networks, by paying special attention to the network architecture. In a second phase, the number of hidden layers is increased. By applying strategies that are particularly suited for deep learning, the results are improved significantly. These strategies include regularization, smarter initialization schemes and more suited activation functions. Experimental results on the TD2003 dataset of the LETOR benchmark show that these well-trained deep neural networks outperform the state-of-the-art algorithms.