IEEE Technology and Society Magazine - December 2016 - 40

Are recommendation systems
ultimately beneficial to society, or do
they potentially exacerbate existing
social problems?

themselves and the way they lead their lives. Since
2008, the Norman Lear Center, a research institute
based at the University of Southern California, has
explored whether it is more effective to understand
audiences in terms of their beliefs and their taste rather than their demographics. In a series of U.S. national
surveys, the Lear Center collected data in order to see
whether there is a correlation between entertainment
preferences, what people enjoy, and political ideology,
what people believe [1].
Survey respondents often have trouble labelling
themselves politically, and so the Lear Center, in
partnership with Zogby Research, created an instrument that would diagnose the survey respondent's
ideology based on their responses to dozens of statements about hot-button political issues. Using
k-means cluster analysis, researchers discovered that
three groups emerged from the national sample. The
most accurate labels for those groups, whose number and size only emerged from the cluster analysis,
were Conservatives, Liberals and Moderates. These
same respondents were asked about their preferred
leisure-time activities and their favorite radio and TV
shows, Web sites, movies, games, news, sports and
much more. Each of these clusters had distinctly different entertainment and leisure preferences: the
only exceptions were football, rock & roll, and the TV
show House, which were equally enjoyed across the
ideological spectrum.
When the clusters were analyzed for demographic
traits, there was a lot of inconsistency. Whites were
more likely to be conservative, blacks liberal. Urban
inhabitants were more liberal; rural area residents were
more conservative. But the ideological typologies demonstrated blurred matrices of similar characteristics on
income, race, gender and residency. There was much
more coherence among each group's taste than there
was in its demographic profile. In other words, demographics could not predict inclusion in ideological
groups but taste preferences could.
A key finding from this research was that advertisers,
marketers, and the media business in general were
doing themselves and audiences a disservice by using

demographics to market to audiences. If they really
wanted to locate their customers, especially with longtail goods, they should be focused on taste, not age,
gender, race, or income.

Popping the Filter Bubble
The Lear Center has been monitoring trends and performing research on social media for several years and
one focus of that research has been on "taste communities" - how social media offers people unprecedented
opportunities to aggregate around their interests and
take control of their own media consumption. Recommendation systems on sites like Amazon, Netflix, YouTube, Spotify, and Pandora have grown steadily more
sophisticated. Investments in these mechanisms appear
to indicate that online businesses are moving away from
demographic marketing to marketing based on customer taste.
But there are concerns about the social problems
that these systems might reinforce. In his popular TED
talk about the danger of "filter bubbles," Eli Pariser [3]
argued that these filters effectively insulate people from
opinions and content that they do not understand or
like. His argument appeared to confirm the existence of
the deeply divided ideological and taste communities
identified in the Lear Center's survey research. It raised
the question: were recommendation systems ultimately
beneficial to society or were they potentially exacerbating existing social problems?
Interviews with people who were working with or
within companies that utilize these systems revealed
some counter-intuitive discoveries. Businesses such
as YouTube were deliberately trying to offer recommendations that were not simply "more of the
same." Often, there are many variations of the same
theme, or literally the same video, on YouTube,
which requires the platform to find a way to suggest
videos that are not virtually identical to the one a
viewer just saw. Visitors may like something similar, but not the same. If an algorithm is optimized
to deliver the best match, the results will probably
be unappealing.
A review of publications from the Association for
Computing Machinery (ACM), revealed several papers
proposing methods for increasing the amount of
diversity, novelty, and serendipity in recommendation engines. The problem is balancing "accuracy" -
the predictive factor determining the likelihood that
an end user will actually like the recommended
content - with newness, difference, and the pleasure
of unexpectedness.
One of the key findings from the politics and entertainment surveys is that a large proportion of the U.S.
population has a predisposition toward things that do

IEEE TEchnology and SocIETy MagazInE

∕

DECEmbEr 2016

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - December 2016