A systematic review of listwise learning to rank (Ba)
Topic for a bachelor/master's thesis
In the literature on preference learning, the problem of "learning to rank" (LTA) has arguably received the most attention so far. Broadly speaking, LTA deals with the task of leveraging suitable training data to learn a “ranker” in the sense of a function that produces predictions in the form of rankings (typically total orders) of query items. Depending on the concrete setting, different types of ranking problems can be distinguished, such as instance, object, and label ranking . Moreover, depending on the type of training data used and the type of loss function minimised by the learner, a distinction between pointwise, pairwise, and listwise learning to rank has been made.
In the listwise approach , the learner seeks to leverage complete rankings (lists) as training information, instead of reducing such information to pairwise comparisons or preference information about individual items. Indeed, such reductions may come with a certain loss of information, which is avoided by the listwise approach. On the other side, listwise learning is challenging and often necessitates various approximations, which may in turn compromise its usefulness. The goal of the thesis is to provide a systematic overview of existing work on listwise learning to rank, along with a critical discussion of its advantages and disadvantages.
The requirements for the thesis include the following: Familiarization with the topics of preference learning and learning to rank; literature search with a focus on listwise LTA; review and systematic exposition of existing work on this topic; critical discussion and comparison with other approaches.
Strong background in machine learning.
-  J. Fürnkranz, E. Hüllermeier. Preference Learning. Springer, 2011.
-  Z. Cao, T. Qin T.Y. Liu, M.F. Tsai, H. Li. Learning to rank: from pairwise approach to listwise approach. Proc. ICML, International Conference on Machine learning, 2007.