Information Technology Reference
In-Depth Information
Chapter 9
Transfer Ranking
Abstract
In this chapter, we introduce transfer learning for ranking, or transfer
ranking for short. Transfer ranking is a task to transfer the knowledge contained
in one learning-to-rank dataset or problem to another learning-to-rank dataset or
problem. This is a typical task in real ranking applications, e.g., transferring from
an old training set to a new one, or transferring from one market to another. In this
chapter, we will briefly review the feature-level transfer ranking approach and the
instance-level transfer ranking approach, and discuss the future research directions
along this line.
Transfer learning is concerned with using a related but slightly different auxiliary
task to help a given learning task.
1
The concept of transfer learning can be applied
to many real ranking scenarios. Here we give some of them as examples.
•
When training a ranking model for a search engine, large amounts of query-
document pairs were previously labeled. After a period of time, these data may
become outdated, since the distribution of queries submitted by users is time-
varying. The prediction on these outdated data can serve as an auxiliary task to
the prediction on the new data.
•
Suppose a ranking model is desired for a newly born vertical search engine while
only labeled data from another vertical search engine are available. Then the pre-
diction on the old vertical search data is an auxiliary task to the prediction on the
new vertical search data.
•
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 mar-
ket. To tackle this problem, many search engines leverage the training data from
the main-stream market to make the ranking model for the niche market more ef-
fective and reliable. It is clear that the prediction on the main-stream market data
is an auxiliary task to the prediction on the niche-market data.
Usually, in transfer learning, the (small amount of) labeled data for the given
learning are is called the “target-domain data”, while the labeled data associated
1
See
http://socrates.acadiau.ca/courses/comp/dsilver/NIPS95_LTL/transfer.workshop.1995.html
.
