Ranking with Deep Neural Networks

The task of ranking is crucial in information retrieval. With the advent of the Big Data age, new challenges have arisen for the field. Deep neural architectures are capable of learning complex functions, and capture the underlying representation of the data more effectively. In this work, ranking is reduced to a classification problem and deep neural architectures are used for this task. A dynamic, pointwise approach is used to learn a ranking function, which outperforms the existing ranking algorithms. We introduce three architectures for the task, our primary objective being to identify architectures which produce good results, and to provide intuitions behind their usefulness. The inputs to the models are hand-crafted features provided in the datasets. The outputs are relevance levels. Further, we also explore the idea as to whether the semantic grouping of hand-crafted features aids deep learning models in our task.