Information Technology Reference
In-Depth Information
Part III
Advanced Topics in Learning to Rank
In this part, we will introduce some advanced learning-to-rank tasks, which are in a
sense orthogonal to the three major approaches to learning to rank. On the one hand,
these tasks consider specific issues that are not core issues in the major approaches.
On the other hand, for any of the three major approaches, the technologies intro-
duced in this part can be used to enhance them or extend them for new scenarios of
learning to rank.
First, we will introduce the task of relational ranking, which explicitly explores
the relationship between documents. This is actually an abstract of many real rank-
ing scenarios. For example, in topic distillation, we will consider the parent-child
hierarchy of webpages in the sitemap, and in diversity search, we would not like to
see many topically similar documents ranked in the top positions. Here the parent-
child hierarchy and topical similarity are both what we call “relationships”.
Second, we will introduce the task of query-dependent ranking. As we can see,
although different kinds of loss functions have been studied, it is usually assumed
that a single ranking function is learned and used to answer all kinds of queries.
Considering that queries might be very different in semantics and user intentions, it
is highly preferred that the ranking function can incorporate query differences.
Third, we will introduce the task of semi-supervised ranking. As one may have
noticed, most of the learning-to-rank work introduced in this topic requires full su-
pervision in order to learn the ranking model. However, it is always costly to get
the fully labeled data, and it is much cheaper to obtain unlabeled queries and docu-
ments. It is therefore interesting and important how to leverage these unlabeled data
to enhance learning to rank.
Last, we will introduce the task of transfer ranking. In practice, we will often
encounter the situation that the training data for one ranking task are insufficient,
but the task is related to another task with sufficient training data. For example,
commercial search engines usually collect a large number of labeled data for their
main-stream market, however, they cannot do the same thing for every niche market.
Then how to leverage the training data from the main-stream market to make the
ranking model for the niche market more effective and reliable is a typical transfer
ranking problem.
After reading this part, the readers are expected to understand the advanced topics
that we introduce, and to be able to formulate other advanced learning-to-rank tasks
based on their understandings of various ranking problems in information retrieval.
