Application Of Ties, Transfer Learning And Pseudo Feedback On Learning To Rank

The ranking function is a core component of information retrieval systems since it directly impacts the relevance and quality of the retrieved results. Recently, learning to rank become popular at the interaction of information retrieval, machine learning, data ming and other related fields. Learning to rank first collects labeled training data and then constructs ranking functions through fitting these training data. In this paper, we propose to improve existing learning to rank methods through solving their issues. Specifically, we propose three new algorithms, addressing the problem of forms of training data, transfer learning and query representation:1. We propose a new type of train data, ties, to complement the traditional preference data.2. We make use of training data from a different task to enhance to learning process of the target task. Thus, we can reduce the amount of labeled training data for the target task.3. Through psuedo feedback, we enhance the query representations for traditional learning to rank methods in order to deal with the diversity of queries effectively.