review articles
Ranking also plays a crucial role in
search engines and recommendation
systems—two prominent data science
applications that we focus on in this
article. Search engines (for example,
Google and Bing) rank Web pages,
images, and other items in response
to a query. Recommendation systems
endorse items by ranking them using
information induced from some context—for example, the Web page a user
is currently browsing, a specific application the user is running on her mobile phone, or the time of day.
In the industrial search and information retrieval community, the
practice of adapting content to satisfy
a search engine’s ranking function or
that of a recommendation engine has
been typically referred to as search
engine optimization (SEO), which has
become one of the most profitable professions in the data science era. Accordingly, there has been much work on devising algorithms that block spammers
and guarantee that content presented
to users is of high quality. Yet, virtually
all retrieval and recommendation algorithms ignore post-ranking/recom-mendation effects (that is, manipulations applied to items) on the corpus.
As it turns out, this reality results in
underoptimized fundamental paradigms for devising search and recommendation algorithms as we discuss
here. For example, a publisher of a
Web page can try and mimic other Web
pages to promote her page in rankings,
thereby potentially causing content
IN HER POPULAR book, Weapons of Math Destruction,
data scientist Cathy O’Neil elegantly describes to the
general population the danger of the data science
revolution in decision making. She describes how
the US News ranking of universities, which orders
universities based on 15 measured properties, created
new dynamics in university behavior, as they adapted
to these measures, ultimately resulting in decreased
social welfare. Unfortunately, the idea that data
science-related algorithms, such as ranking, cause
changes in behavior, and that this dynamic may lead
to socially inferior outcomes, is dominant in our new
online economy.
Rethinking Search
Engines and
Recommendation
Systems:
A Game Theoretic
Perspective
DOI: 10.1145/3340922
Novel approaches draw on the strength
of game theoretic mechanism design.
BY MOSHE TENNENHOLTZ AND OREN KURLAND
key insights
˽ An overall modeling and analysis
of the search and recommendation
systems ecosystems (publishers, users,
mediators, and competitors) is essential
for their efficient design.
˽ Strategic content dynamics may lead
to failure of classical principles of
search and recommendation systems
in maximizing social welfare; these
principles can/should be revisited.
˽ Game-theoretic mechanism design
can be adapted to the design of
economically efficient search and
recommendation systems.