follow a Bell curve distribution, scale- free networks follow a power law distribution,where the big amount of nodes with very few connections shape the long tail.The concept of the long tail is described in
Clay Shirky’s essay “Power Laws,Weblogs, and Inequality” (200
6), where the internet theorist notes that a small amount
of weblogs have a huge amount of connections leading to them, while a “long tail” of
the vast majority of weblogs have very few links (41). An important point that needsto be emphasized is how the distribution of scale-free networks came to be. Accordingto Barabasi and Bonabeau, when people are choosing where to link to, they seem to
have a “preferential attachment” towards nodes which already have a large amount of
connections (2003: 54-55). Clay Shirky also attributes this scale-free distribution to a
freedom of choice: “[t]he very act of choosing, spread widely enough and freelyenough, creates a power law distribution” (2006: 36).
In analyzing the Youtuberecommendation system, I will argue that the recent development of its algorithmfunctions as an additional factor, together with the factor of the freedom of choice, inthe formation of power law distributions.
In their paper “The Youtube Video Recommendation System” (2010), Google
software engineers James Davidson et al. explain that the recommendation feature of
Youtube is based on the user’s activity on the website, which consists of watching,
nd sharing videos. Those videos function as “seeds” for the formation of a setof “similar videos”, that the user would be likely to watch after the seed videos (294).The authors refer to the main factors which define this “likelihood” that the
recommended videos will be appreciated by the user. Here, two points will beemphasized on: first, the algorithm of the recommendation system considers the viewcount and ratings of a video from previous users, the activities of commenting,
“favoriting” and sharing, as well as whether the video has been watched in its entirety
(295). Second, Davidson et al. mention that the aforementioned process is facingcertain limits, for example, some videos have a very low view count and, therefore,
the algorithm is not “
able to compute a reliable set of related videos
” for them (294).
As it becomes apparent from the description of this process, the Youtube videorecommendation system for an individual user is based on data retrieved by the
choices of previous users. Connecting this information to Shirky’s “freedom of choice” on the web, we can
observe the formation of an unequal distribution in thenetwork of Youtube videos. A high view count from previous users results in moreconnections to recommended videos, while videos with a very low view count are nottaken into consideration from the system and, thus, are not connected to other videosas recommendations. This process intensifies the power law distribution, drawing the
long tail further away from the “head” of the few most popular videos
. As Barabasi
and Bonabeau put it “this "rich get richer" process will gene
rally favor the early
nodes” (55).As this essay has attempted to show, the mechanism of “preferentialattachment”
in Youtube is reinforced by the recent development in
personalized recommendation system. It could be argued that recommendationpersonalization, a significant development of the Web 2.0 online experience, shrinks
the user’s online space, limiting the content they have the chance to stumble upon.
This finding highlights the importance of the awareness of how recommendationalgorithms function, as it could enlighten us on how democratic theserecommendation systems are.