algorithmic-extremism.html.pm

#lang pollen

◊define-meta[page-title]{A critique of "Algorithmic Extremism"}
◊define-meta[short-title]{Critique: "Algorithmic Extremism"}
◊define-meta[original-date]{2020-01-19}
◊define-meta[snippet]{A critique of a recent observation of YouTube recommendations}

This post is a critique and recontextualization of a recent pre-print
paper by Mark Ledwich and Anna Zaitsev, "[Algorithmic Extremism:
Examining YouTube's Rabbit Hold of
Radicalization](https://arxiv.org/abs/1912.11211)."

First, an alternative abstract:

◊q{

The role that YouTube and its behind-the-scenes recommendation
algorithm plays in encouraging online radicalization has been
suggested by both journalists and academics alike. This study
quantifies the extent to which this is true for users who are new to
the platform and who do not have a viewing history. After categorizing
nearly 800 political channels, we were able to differentiate between
political schemas in order to analyze the algorithm traffic flows out
and between each group. We analyzed the recommendations that an
anonymous user was provided when visiting each channel type. We
observed that YouTube's recommendation algorithm actively discourages
new and anonymous viewers from visiting radicalizing or extremist
content. Instead, in this initial exploratory phase, the
recommendation algorithm favors mainstream media and cable news
content over independent YouTube channels. Our study thus suggests
that YouTube's recommendation algorithm does not promote inflammatory
or radicalized content to users who are new to the platform.

}

That is how I would have framed this paper.

Insights into the black box of any proprietary recommendation
algorithm are hard to come by. Ledwich and Zaitsev provide a valuable
contribution. While YouTube has publicized the technical details of
their recommendation algorithms, they publish scant data on the nature
of the resulting recommendations.

My critique of this paper is that Ledwich and Zaitsev don't adequately
confine their conclusions to the experience of a new user. They use
language that implies that their observations likely generalize to the
recommendation algorithm as a whole:

◊itemize{

"YouTube's recommendation algorithm fails to promote inflammatory or
radicalized content"

"YouTube's recommendation algorithm actively discourages viewers from
visiting radicalizing or extremist content"

"The data shows that YouTube does the exact opposite of the
radicalization claims"

}

They include only one paragraph that expresses a very limited doubt
about whether their observations would generalize to logged-in users
(emphasis mine):

◊q{

One should note that the recommendations list provided to a user who
has an account and who is logged into YouTube might differ from the
list presented to this anonymous account. ◊b{However, we do not
believe that there is a drastic difference in the behavior of the
algorithm.} Our confidence in the similarity is due to the description
of the algorithm provided by the developers of the YouTube algorithm
[[38]](https://daiwk.github.io/assets/youtube-multitask.pdf). It would
seem counter-intuitive for YouTube to apply vastly different criteria
for anonymous users and users who are logged into their accounts,
especially considering how complex creating such a recommendation
algorithm is in the first place.

}

That paragraph extremely understates the role that personalization
plays in YouTube's recommendation algorithm. I recognize that Ledwich
and Zaitsev haven't collected the data needed to confirm themselves
that personalized recommendations are different than anonymous
recommendation, so perhaps they are just being careful. But the whole
point of a recommendation algorithm is to tailor the recommendations
to the individual. And further, the very description of the algorithm
provided by YouTube ([Recommending What Video to Watch
Next](https://daiwk.github.io/assets/youtube-multitask.pdf))
explicitly mentions personalized recommendations that take into
account a user's watch history, demographics, time, and location.

◊fig[#:src "assets/zhao-figure-1.png"]{Figure 1 from Zhao et al.'s
[Recommending What Video to Watch Next: A Multitask Ranking
System](https://daiwk.github.io/assets/youtube-multitask.pdf).}

It's also probably not correct to refer to "the YouTube algorithm" as
having been described in an individual paper. YouTube is constantly
running experiments. Some improvements are published. Some but not all
of those end up in the production algorithm.

So, let's look at another paper from YouTube: [Deep Neural Networks
for YouTube
Recommendations](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf). This
paper also acknowledges that a user's activity history is relevant for
recommendations.

◊fig[#:src "assets/covington-figure-2.png"]{Figure 2 from Covington,
Adams & Sargin's [Deep Neural Networks for YouTube
Recommendations](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf).}

So again, Ledwich and Zaitsev have made some interesting and important
observations about the behaviour of YouTube's algorithm during its
initial, exploratory, un-personalized phase for a new user. But these
observations should not be taken to demonstrate anything about the
behaviour of the algorithm (or the behaviour of the algorithm mixed
with user responses to that algorithm) as it learns more and more
about you.