Query-specific optimal convolutional neural ranker

In this paper, we propose a novel learning-to-rank method by developing a convolutional neural network (CNN)-based ranking score estimation function (ranker). We use the query, query-specific preference, and the neighborhood structure to regularize the learning of the CNN ranker parameters. We propose to impose the CNN outputs of a query-preferred data object to be larger than that of a data object which the query tries to avoid. Also we hope the ranking scores of the data objects can be smooth over neighborhoods and the ranking score of the query itself can be large. We construct a joint unified minimization problem by combining these regularization problems to learn the parameters of CNN, and develop an iterative algorithm based on fix-point method. The experiments over the benchmark data sets of image retrieval and ship roll motion prediction show its effectiveness.