Learning Ranking Function via Relevance Propagation

In this paper, we propose a novel ranking function learning framework based on relevance propagation. The propagation process is used to propagate the relevance scores from labeled documents to other unlabeled ones so that more training data are available to learn the ranking function. It is realized by the manifold ranking algorithm, which has been proved to be very effective in content-based image retrieval. We investigate two kinds of propagation schemes: intra-query propagation, which confines the propagation within the documents returned by a single query, and inter-query propagation, which allows the propagation across the documents returned by multiple queries. Inter-query propagation scheme is proved to be more effective but less efficient since the number of documents that need to be processed are huge. Thus, the intra-query propagation scheme is practically more useful and adopted in this paper. The proposed framework has two main advantages. On one hand, it can effectively utilize the unlabeled data to proliferate the training set. On the other hand, the relevance propagation process is totally independent on the learning algorithms. This property lends the framework remarkable flexibility to leverage any state-of-the-art ranking function learning algorithm. Any improvement made to either propagation or learning can improve the overall performance. Experimental results on commercial search engine data show that the framework attains a significant improvement over existing ranking function learning algorithms.