Information Technology Reference
In-Depth Information
With the above feature representation, three learning-to-rank methods are exam-
ined: Ranking SVM [
7
,
8
], support vector regression, and gradient boosted decision
tree. The first one is a pairwise ranking algorithm, while the next two are point-
wise ranking methods. The TREC 2002, 2003, and 2004 novelty track data are
used as the experimental datasets. R-Precision is used as the evaluation measure.
The experimental results basically show that support vector regression and gradient
boosted decision tree significantly outperform a simple language modeling baseline
and Ranking SVM. Furthermore, gradient boosted decision tree is very robust and
achieves strong effectiveness across three datasets of varying characteristics.
The above experimental findings are different from the experimental findings
on other tasks such as document retrieval and question answering. This is possibly
because sentence selection for summarization is by nature more like a classification
problem for which the pointwise approach can work very effectively, and even more
effectively than the pairwise approach.
14.5 Online Advertising
The practice of sponsored search, where the paid advertisements appear alongside
web search results, is now one of the largest sources of revenue for search engine
companies.
When a user types a query, the search engine delivers a list of ads that are relevant
to the query adjacent to or above the search results. When a user clicks on an ad,
he/she is taken to the landing page of the advertisement. Such click generates a
fixed amount of revenue to search engine. Thus, the total revenue generated by a
particular ad to the search engine is the number of clicks multiplied by the cost per
click.
It is clear that the search engine will receive more clicks if better matched ads are
delivered. Thus, a high accuracy of delivering the most preferred ads to each user
will help the search engine to maximize the number of clicks. However, this is not
all that the search engines want. The ultimate goal of a search engine is to find an
optimal ranking scheme which can maximize the total revenue. Specifically, besides
the number of clicks, an extremely important factor, which substantially influences
the total revenue, is the bid price of each ad. Therefore, the best choice for defining
the ad ranking function is to consider both the likelihood of being clicked and the bid
price of an ad (i.e., the maximization of the revenue). There are several works that
apply learning-to-rank technologies to solve the problem of revenue maximization,
such as [
5
,
9
,
19
]. Here we will take [
19
] as an example for detailed illustration.
Once again, the first step is to extract features to represent an ad. Specifically, in
[
19
], the following features are extracted.
•
Relevance features, including term frequency, TF-IDF, edit distance, BM25 of
title, BM25 of description, LMIR of title, and LMIR of description.
•
Click-through features, including ad click-through rate (CTR), campaign CTR,
and account CTR.
•
Other features, including bid price and match type.
