A Network View of Social Media Platform History: Social Structure,
`Dynamics and Content on YouTube
`
`John C. Paolillo
`Indiana University Bloomington
` paolillo@indiana.edu
`
`Sharad Ghule
`Indiana University Bloomington
` sharadsghule@gmail.com
`
`Brian P. Harper
`Indiana University Bloomington
` bpharper@indiana.edu
`
`Abstract
`
`Social media sites are prone to change from many
`internal and external causes, yet it is difficult to
`directly explore their histories in terms of the content
`itself. Search and browsing features are biased toward
`new and paid content, archives are difficult to navigate
`systematically, and their scale makes any observations
`challenging to contextualize. Here, we present results
`of an ongoing study of YouTube’s history (currently
`with more than 76 million videos) using a combination
`of iterative browsing, network crawling and clustering
`within and across time periods. Through this method,
`we are able to identify historical patterns in YouTube's
`content related to internal and external events. Our
`approach thus illustrates an adaptation of network
`analysis for understanding the content histories of
`social media platforms.
`
`1. Introduction
`
`Currently, YouTube is at a crossroads: YouTube’s
`dominance in online video is now challenged by
`Amazon, Facebook, Hulu, Netflix and Twitch.
`YouTube’s visibility has exposed it to regulatory
`scrutiny and advertiser protests, threatening revenue. In
`response, YouTube has changed
`its advertising
`algorithms and upset the economic viability of many
`channels, alienating channel owners. Any of these
`conditions could induce large changes on the site,
`shaping its content or what we can access of it.
`We therefore need a history that would chronicle
`the emergence and
`influence of
`the platform's
`dominant genres and content types since 2005, ideally
`indexed to changes in the platform's features and
`incentives as well as external world and media events.
`YouTube has archival properties, however, and the
`YouTube public data API reflects
`the historical
`character of the site through the publication dates of
`video and channel metadata. Channels and their videos
`are also structured as a network, via relations such as
`liking and favoriting videos. Can this information be
`used to further illuminate the history of the site?
`
`Our answer to this question is yes, based on a
`network analysis in which the publication dates of
`videos are used to segment the YouTube network into
`a sequence of time slices, covering its entire history
`from May 2005 to December 2016. This analysis
`reveals the evolution of a range of different genres of
`content, which can be read in terms of responses to
`historical events and platform changes. This work
`provides a potentially
`important frame for
`the
`interpretation of past and current studies of YouTube
`content.
`
`2. Literature Review
`
`From its initial pre-launch public availability in
`2005, YouTube rapidly became the dominant platform
`for the distribution of online video. This 12-year
`history has been unstable, punctuated by technical
`changes
`to
`the platform, purchase by Google,
`introduction of advertising, international expansion, for
`example. External events have also had effects: large
`user migrations, political events, copyright lawsuits,
`changes in national and international regulation of
`internet technology, major studio participation in
`YouTube, and the US presidential elections have all
`been felt in different ways by YouTube users.
`Empirical research
`insufficiently contextualizes
`YouTube’s content and its evolution. Early attempts at
`a global-scale analysis of YouTube’s content exist [1],
`but they are either small in comparison to its actual
`scale at the time [2], are based on specific events [3],
`or they do little to address the nature of the content or
`how it might relate to platform features [4, 5]. A
`representative compilation of early
`research on
`YouTube is The YouTube Reader [6]. Early histories of
`the platform exist [7], but numerous changes in the site
`have obscured the relationships among YouTube’s
`features, users, content and external events.
`Other YouTube research has addressed YouTube’s
`politics as a platform [8], the recommender system [9,
`10, 11], social network effects on content propagation
`[3, 12, 13], the features of memes [14], multichannel
`
`Proceedings of the 52nd Hawaii International Conference on System Sciences | 2019
`
`URI: https://hdl.handle.net/10125/59701
`ISBN: 978-0-9981331-2-6
`(CC BY-NC-ND 4.0)
`
`Page 2632
`
`Comcast, Ex. 1143
`
`

`

`
`
`networks [15], and even specific genres of content
`[16]. These pieces often exist
`in
`isolation of
`YouTube’s development over time, as can be seen in
`the contradictory findings at different times regarding
`the popularity of longer videos [17, 18].
`An important contextual component missing from
`the discussion of YouTube is the role of mutual
`support among channels in the cultivation of its genres.
`YouTube’s liked and favorited video playlists offer one
`record of such support, which also flows and ebbs over
`time, as channels become active or dormant. Such
`social processes have been shown to be instrumental in
`genre emergence [19], and a network analysis offers
`one approach for revealing them [20]. Time in network
`analyses, however, has no standardized treatment. We
`therefore ask: how can we use the network of likes and
`favorites among channels to read a history of genre
`evolution on YouTube?
`
`3. Method
`
`The method employed in this study has three main
`components: (i) construction of a sample using
`browsing and crawling and the Google/YouTube
`public data API, (ii) extraction of time-located network
`samples and clustering them, and (iii) organizing and
`interpreting the timeline of network clusters. Each of
`these corresponded to three distinct phases of research,
`discussed in turn below.
`
`3.1 Sampling YouTube
`
`YouTube is large and unwieldy, and its complete
`data are accessible only within Google. Data for
`individual videos are exposed only through search and
`browsing functions that are subject to unknown biases
`(e.g., sponsored search
`features and
`the video
`recommendation algorithm) and cannot be sampled in
`a truly random manner, and we must resort to crawling
`a large sample. Problematically, crawled samples miss
`unconnected components. Consequently, a diversified
`strategy for sampling is necessary, relying on searching
`and browsing to identify starting points for crawling,
`and iterative phases of both activities.
`The initial sample for this study was based on a
`collection of YouTube channel IDs identified for a
`project on conspiracy theory videos in 2015, using
`searching and browsing strategies. A script written for
`the Firefox Greasemonkey plugin was used to collect
`channel IDs into a PostgreSQL database directly while
`browsing. In addition, the script reports whether the
`channel for the current page was already recorded in
`
`the database. YouTube search was used to initiate
`browsing, and browsing strategies were developed so
`as to rapidly gather distinct channel IDs. On a typical
`video page, the first video listed on the right bar often
`comes from the same channel, and the second is an
`advertisement. Videos from the third on come from a
`range of channels: the same channel, related channels
`and "recommended" channels. The last of these are fed
`by a YouTube algorithm that references a user’s
`viewing history; typically these have already been
`visited. We therefore focused attention on videos after
`the first two with unfamiliar channel names, using the
`thumbnails and titles to help recognize if a particular
`video had already been seen. When the initial project
`was broadened beyond conspiracy theories, the same
`strategies were employed, merely using different
`YouTube searches from which to begin browsing.
`Channel IDs from browsing became the seed set for a
`crawl collected through the YouTube public data API.
`Each channel
`is associated with
`three playlists:
`uploads, likes and favorites. The first is merely the list
`of the videos uploaded by the channel; the second and
`third represent videos that users have identified as ones
`they
`like or favorite, using YouTube's
`interface
`features. Typically, these videos are ones produced by
`other channels (though they need not be). YouTube's
`recommendations are generated partly from videos that
`are co-liked or co-favorited with the video being
`watched. Hence, crawling these two playlists to obtain
`the video information and that of their associated
`channels tends to expand the set of channels observed
`while mirroring YouTube's video recommendations.
`Unfortunately, crawling via the API has limitations.
`It does not list channels that liked or favorited a
`particular video, so we must always identify channels
`first. This requires that all our channels post videos,
`when many do not. Relations to such profiles could be
`crawled through the comments feature, but this would
`expand the data collection beyond the capabilities of
`our current system architecture. Similarly, channel
`subscriptions are treated as private by the API, and for
`non-posting channels, likes and favorites can also be
`made private. Without appropriate searching and
`browsing strategies it is likely that sections of the
`network would be missed, especially less popular
`channels. For this reason, the searching/browsing and
`crawling processes were repeated several times from
`July 2015 to March 2017, ending with a sample of
`76,081,372 videos and 549,383 channels.
`The resulting database contains metadata for a
`small but popular and highly connected fraction of the
`total activity on YouTube. Although our sampling
`began with the conspiracy theory channels, these are a
`
`Page 2633
`
`

`

`
`
`small proportion of the final network, which is
`otherwise dominated by entertainment content (below).
`
`3.2 The Network Over Time
`
`Our network analysis of YouTube is based on the
`structure induced by likes and favorites; we treat these
`as indicating directed links between channels, i.e., a
`channel has a (directed) link to another channel as
`strong as the number of times the first channel likes or
`favorites videos uploaded by the second. We treat likes
`and favorites as equivalent because the two relations
`are strongly correlated [4]. Likes and favorites also
`tend to occur in a short window of time after a video is
`released [4]. For this reason, we use the video
`publication date of the liked/favorited video as a proxy
`for historically dating the relationship.
`Using 3-month intervals over the video publication
`dates as a moving window in which to examine
`connectivity of channels, we segmented the network
`into 141 samples, starting from April 2005, shifting the
`window by one month for each sample, and ending
`with the December 2016 sample. To keep our networks
`within a size that we could process, we used a
`threshold of a minimum of 10 likes/favorites from one
`channel to another within any given sample to include
`a link in the network.
`
`3.3 Clustering
`
`There are many approaches to clustering networks
`[21]; here, we employ the Louvain method of [22].
`This algorithm performs well for large networks,
`especially with a high clustering coefficient and a fat-
`tailed degree distribution, as occurs in the YouTube
`network [4]. It performs an agglomerative hierarchical
`clustering in which a node is assigned to a cluster if
`doing so maximizes the modularity of the network,
`continuing until either a single node remains or
`modularity cannot be increased further. Modularity
`clustering is not perfect: it sometimes infers non-
`existing relationships between clusters based on weak
`false positive links [21], and tends to give large
`numbers of clusters in sparse networks. Nonetheless, it
`works well for detecting well-defined but small
`clusters in very large networks, as we expect to be the
`case with YouTube. Since crawling biases our samples
`toward connectivity, we anticipate some issues with
`interpretability in larger clusters.
`Compatible implementations of this algorithm exist
`in Gephi [23] as "modularity class", and in the igraph
`package [24] as the function cluster_louvain(). For
`clustering the samples, we use the implementation in R
`
`[25]. This results in anywhere from 1 to 3747 clusters
`for each sample depending on its size and overall
`connectivity. Clusters are identified by arbitrary ID
`numbers and the only means for identifying them
`across different samples is through their aggregate
`common memberships, which we obtain by cross-
`tabulating clusters from successive pairs of samples.
`This is identical to a treating the entire sample as a
`network of clusters, in which links represent shared
`membership between clusters across different samples.
`For convenience, the cluster comparison network
`was imported in its entirety into Gephi as a directed
`network with no minimum value for a link. Using
`Gephi's modularity class, we assigned each of the
`clusters within samples to new cross-sample clusters.
`Clusters with substantial overlap or that regularly
`exchange members fall together into a single new
`cluster assignment; clusters whose membership largely
`excludes those of another cluster over time appear in
`distinct new clusters, thereby identifying clusters with
`stable yet evolving membership across time. The
`success of this approach depends on the suitability of
`the threshold for the initial samples, the size of the
`moving sample window, the frequency of the samples,
`and the stability of class membership over time, so
`changes in these values would yield different results.
`Gephi also provides network layouts; a suitable
`layout for this data should be able to find a linear
`structure or structures, showing the evolution and
`relative closeness of different content clusters. We
`used two force-directed layouts: the Yifan Hu layout
`for rapidly finding the global structure, and Force Atlas
`to verify that the observed structures were not peculiar
`to Yifan Hu.
`The resulting layout appears as Figure 1, in which
`we find a single linear structure whose two large bends
`and single sharp elbow correspond to gradual and
`sharp changes in cluster membership, respectively. The
`layout has been rotated so that clusters from the earliest
`samples are on the left, and tracing along the main
`connected path takes one through more recent samples,
`to the final sample on the far right. Nodes in Figure 1
`represent the sample modularity classes, with color
`indicating the cluster a node belongs to and size its
`number of members.
`The largest 24 of the clusters (out of 14978 total)
`account for 75.9% of the network’s nodes, with the
`next largest containing only 0.1%. Individual clusters
`are rendered in Figure 2, so that their lifespans can be
`more readily recognized, alongside their relative sizes
`and general type of content (this indicates in which
`subsection it will be discussed below). The largest
`nodes group around a central path, with fine filaments
`
`Page 2634
`
`

`

`
`
`representing the paths through the smaller classes and
`clusters extending outward from it on either side,
`including smaller clusters not shown in Figure 2, as the
`full layout exceeds the margins of the image. To clarify
`the cluster timelines, we produced Figure 3, in which
`each cluster is represented by a horizontal bar spanning
`the x-axis from its beginning point to its endpoint.
`Scanning vertically in Figure 3 indicates which clusters
`overlap at specific times.
`To facilitate cluster interpretation, we created a web
`interface that provided summaries of the number of
`videos in each cluster for each month of the sample,
`along with a listing of ten videos from each of the 100
`most connected channels in the cluster and active links
`to the videos and channels on YouTube. All three co-
`authors explored the full complement of clusters
`through this interface, meeting together to discuss and
`reconsider their interpretations.
`
`4. Interpretation of the network
`
`
`A few observations can be made from Figures 1
`and 2 directly. First, there is a single central core to
`YouTube’s network with varied content, as reported in
`[2]; it is stable over YouTube’s history, although its
`composition changes. Many filaments diverge from the
`core, carrying channels toward or away from it, but
`
`these account for only a quarter of the observed
`network. In other words, YouTube's content is not
`strongly segmented due to, e.g., language markets,
`political polarization or content, as might have been
`expected. Such a pattern would appear as multiple,
`disentangled paths in the network, arising from clusters
`whose exchange of members with the core clusters is
`less frequent. We turn now to the specific patterns of
`content within the clusters that can be observed.
`
`4.1 The Early Years of YouTube
`
`
`Clusters 387, 629, and 909 represent the first stages
`of the development of YouTube's content. Cluster 387
`arises in October 2005, just 8 months after YouTube
`became public;
`it contains mostly music-based
`channels, typically songs made by YouTube users or
`remixes of popular songs. This cluster also contains
`viral videos (for example, “Charlie bit my finger -
`again,” “Evolution of Dance,” and “Sneezing Baby
`Panda”), indicating their importance in YouTube’s
`early history (they are otherwise infrequent). Cluster
`387’s
`content
`is predominantly
`entertainment,
`suggesting that the platform served a limited function
`in its early phase. Clusters 629 and 909 branch off
`from 387, maintaining continuity in both having music
`channels.
`
`Figure 1. Final layout of network of shared membership in modularity classes of 141 sample networks,
`based on 3-month samples of YouTube channel-to-channel likes and favorites spaced at overlapping one-
`month intervals.
`
`
`
`Page 2635
`
`

`

`
`
`Figure 2. Clusters of modularity classes in the network of Figure 1, grouped in columns by type of
`content. Size of each cluster as percentage of modularity classes of the cluster comparison network is
`given in the lower right corner of each panel.
`
`
`Figure 3. Temporal relationships of clusters in Figure 1, grouped by type of content. The left-right location
`and extent of a bar indicates the time period occupied by a cluster; clusters that overlap vertically occur at
`the same time.
`
`Page 2636
`
`

`

`
`
`Cluster 629 features more channels with hip-hop
`and non-pop music genres, alongside other early news
`channels such as the BBC News channel with material
`from its regular broadcasts, and The Young Turks, a
`news/commentary channel created specifically for
`Internet-based consumption. Both 387 and 629 end
`while increasingly accruing content related to the 2008
`US presidential election. The official channel for the
`Barack Obama campaign especially gains prominence
`in summer 2008 in cluster 629, but both clusters
`feature electoral content.
`Cluster 909 features music, comedy (especially for
`young males), and early YouTube-specific channels
`like Machinima, College Humor, and Rooster Teeth. It
`persists later than 387 and 629, possibly owing to its
`lack of political content from the 2008 election. Like
`cluster 629, broadcast media sources appear in 909, but
`while the media in 629 was political, 909 contains
`entertainment, such as the BBC cars and comedy show
`Top Gear. Cluster 909’s last months overlap with the
`launch of Vevo channels, which likely served to
`disrupt the music channels present in this cluster.
`
`4.2 The Two Main Streams
`
`The differences in 629 and 909 become amplified
`after this point, yielding two distinct streams of
`subsequent clusters lasting through the rest of the
`timeline: one with primarily entertainment content and
`the other with political content. The entertainment
`stream, originating from cluster 909, is larger and more
`fluid, with some members separating into specific
`subculture
`topic clusters and
`rejoining popular
`entertainment clusters
`later. The politics stream,
`partially descended from 629, is more stable, and while
`it possesses multiple clusters within it, they have strong
`temporal continuity from cluster 629 onwards. Clusters
`outside these two streams represent language markets
`outside the dominant English-language market on
`YouTube.
`
`4.3 The Entertainment Stream
`
`The Entertainment stream is broad, with the three
`largest clusters in the network. Its size and diversity
`make it somewhat difficult to characterize, but several
`consistent elements recur: music, video games, and
`comedy. This stream begins in late 2009 with clusters
`3327 and 2615, around the time Viacom’s settlement
`with Google that resulted in the Content ID system.
`Google simultaneously changed its rating system from
`a five-star rating system to merely reporting likes. Both
`changes arguably affect
`the connectivity among
`
`channels, and hence may play some role in explaining
`the emergence of two distinct clusters. Cluster 2615 is
`dominated by copyright-protected popular music in the
`form of Vevo channels, i.e., predominantly pop and
`country music. There are occasional parkour and
`skateboarding channels, whose videos often contain
`musical backing similar to Vevo's pop music, though
`they are clearly not professionally mixed. Cluster 3327
`contains music and content created within
`the
`YouTube platform itself, outside of any arrangement
`with Vevo. Its music comes predominantly from
`electronic genres such as dubstep and techno. Cluster
`3327 also shows some of the early video bloggers
`(“vloggers”).
`Cluster 4898 wraps around the elbow in our data
`representation, and both clusters 2615 and 3327 feed
`into it, though not at the same time. This period
`appears to correspond to the shift in YouTube’s
`definition of a view to include watch time in March,
`2012. Prior to the elbow and the sudden decline of
`cluster 3327, 4898 is characterized by a mixture of
`video games and music content. The gaming content is
`not centered clearly on a specific game or genre, and
`the music content is partially Vevo channels formerly
`in cluster 2615 and partially variations of pop music.
`Lindsey Stirling's channel exemplifies this music; at
`this time she produced violin pieces inspired by video
`game music. After channels from cluster 3327 merge
`into 4898, it becomes dominated by personality-driven
`YouTubers:
`GennaMarbles,
`PewDiePie,
`CaptainSparkles, among others. These channels
`combine some contemporary topics, lifestyle, gaming,
`fashion, etc., with comedy focusing on the host's
`personality, and become central to the entertainment
`stream from this period onwards.
`Rather than remaining a new large cluster, 4898
`splinters into multiple clusters: 6445, 9833, and 8778.
`These show similar divisions as 2615 and 3327 did in
`that while all feature popular culture, 8778, 9833, and
`2615 feature content produced for non-YouTube
`audiences (e.g., broadcast and cable television), while
`6445 and 3327 feature content produced specifically
`for YouTube. Cluster 6445, like 4898 before it,
`features personality-driven YouTubers, primarily in
`gaming
`and
`fashion. Cluster
`8778
`features
`predominantly Vevo and other music channels,
`alongside late night comedy show clips. Cluster 9833
`also features late night comedy channels like The
`Tonight Show Starring Jimmy Fallon, although much
`of the cluster focuses on reviews and critique of
`movies. Sketch comedy channels are also present. The
`presence of musical guests on late night comedy shows
`
`Page 2637
`
`

`

`
`
`helps to explain the presence of these shows in both
`clusters.
`These three clusters feed into the two clusters of
`11844 and 14486 at the end of our sample period.
`Much like cluster 4898, which straddled a major
`change in our network, 11844 begins with general
`gaming content, primarily of Let’s Plays of varied
`games, though Minecraft makes a strong and consistent
`showing. As cluster 6445 ends, many of the channels
`from that cluster shift into 11844, leading to the
`personality-driven channels dominating it. Cluster
`14486 began as a combination of fitness and lifestyle
`channels; at this time it also contains many popular
`videos in Italian, German, and Czech, though often
`with parts of their titles in English or subtitles in
`English. Later, language-specific clusters form around
`some of these channels.
`By mid-2015, after clusters 9833 and 8778 end and
`their constituent channels merge into 11844 and 14486,
`it becomes increasingly difficult to accurately describe
`11844 and 14486, with both clusters becoming so
`broad in scope that, while they are still clearly
`entertainment-focused,
`further characterization
`is
`difficult. This trend increases following the end of
`cluster 11844 in mid-2016 as most of the entertainment
`channel fall into 14486. One possible explanation for
`the difficulties in the period after mid-2015 is the
`candidacy announcements of the 2016 US Presidential
`Election, whose use of late-night comedy TV as a
`platform may have destabilized the entertainment
`clusters. A second is that the implementation of
`YouTube's paid subscription service YouTube Red
`may have disrupted some clusters by putting
`interactions behind a paywall where they are no longer
`observable. A third potential explanation is that the
`recency effect in likes and favorites distorts the
`network connectivity at the end of our sample period,
`closer to the point of our actual observations. The
`precise reason cannot be known right now, but we note
`that the clustering may be lower quality toward the end
`of our sample period.
`
`4.4 Subcultural Entertainment
`
`In addition to the main entertainment structure,
`there are several offshoot clusters of entertainment
`videos throughout the network. These clusters usually
`exist for brief periods of time, after which the channels
`involved rejoin one of the larger clusters of the
`entertainment stream. This is not universally true, and
`some channels
`in subcultural clusters
`regularly
`fluctuate between the main cluster and the subcultural
`
`cluster. Some are also more naturally associated with
`the politics stream rather than entertainment.
`Cluster 7047 is one such subculture cluster, and
`lasts for an unusually long time for this network where
`the constituent channels join either clusters 11844 or
`14486. Originally having short skateboarding and
`parkour videos, longer videos made using GoPro or
`other active point-of-view cameras start to show up in
`7047, allowing an “urban spelunking” genre in which
`the content creators explore rooftops, climb
`tall
`buildings, and generally trespass, often without safety
`equipment. Skateboarding’s
`importance diminishes
`after the emergence of the GoPro videos.
`Another subcultural cluster, 5876, features plane
`flights and train operation as its primary material. Its
`focus is clearly technical, especially the inner workings
`of machines. At different points, this cluster attracts
`channels from the Maker movement (electronic gadget
`and computer hackers) and car tinkerers. It transitions
`into cluster 11709 with relatively little difference in the
`channels of interest, although at that point drones
`emerge as a video platform. Cluster 11709 remains
`distinct from the rest of the entertainment cluster until
`the end of our data collection rather than joining cluster
`11844 or 14486.
`Cluster 7953 is dominated by hyper-masculine
`channels aimed at high school and college-aged males.
`The videos
`in
`this cluster
`include sports, male
`heteronormative dating advice, pranks, and videos of
`gameplay in Call of Duty. More so than among the
`other clusters, videos of these channels tend to feature
`video thumbnails with attractive, scantily-clad women
`on them. This cluster could be characterized as
`appealing to “Bro” culture. A different male-audience
`cluster is 9126, featuring cars, guns, and gadgets,
`which partially overlaps with transportation clusters
`5876 and 11709, although 9126 is favors experiences
`of speed and danger with cars more than their technical
`workings. This cluster also includes gun enthusiasts
`and maker channels. Many videos in 9126 demonstrate
`some technology a channel is dedicated to, thereby
`serving an advertising function. Members of this
`cluster arrive from both the entertainment and political
`(see below) streams, and some channels display salient
`political attitudes alongside the subjects of their
`technical interests, a stance less common in other
`entertainment clusters.
`One subcultural cluster in our network is not
`predominantly masculine in character, namely cluster
`10741. This cluster features relaxing content, primarily
`soft music, nature sounds, and videos of animals. As
`the content appears to be based more on a mood than
`on a particular language, this cluster also has some
`
`Page 2638
`
`

`

`television news media
`Broadcast and cable
`YouTube channels appear in this stream, including
`cable news channels ABC, CBC, CBS, CNN, Fox
`News, MSNBC, and NBC, several local affiliate
`channels of these networks. State-supported networks
`BBC and CBC also appear. However, most of these
`television media outlets arrive in this stream later and
`are not as well integrated into the network as the media
`entities discussed previously. While they spend more
`time in the clusters of the politics stream than
`elsewhere, they often cross over into other clusters
`more focused on entertainment. Despite the popularity
`of
`their
`content outside of YouTube,
`their
`likes/favorites pale in comparison to those of RT. It is
`likely that this represents a deliberate strategy of
`linking on the part of RT (and a reciprocal lack of that
`strategy on the part of other news organizations) rather
`than an absolute index of popularity. In other words,
`different organizations differ in their employment of
`search engine optimization
`techniques, and
`this
`influences the likes/favorites we have observed.
`
`4.6 Specific Language Markets
`
`In addition to these two main streams, we identify
`several major clusters of YouTube channels specific to
`different language markets. Clusters 11151, 11841, and
`12760 represent Spanish, Brazilian Portuguese, and
`Russian
`language channels respectively, and are
`remarkably similar in their positions in the network.
`All three start in mid-2011 and persist until the end of
`our collection in 2016; these three are the longest-lived
`clusters in our network. While the Brazilian Portuguese
`and Russian clusters are heavily identified with their
`country of origin, the Spanish cluster, 11151, is
`somewhat more complex: many of
`the popular
`channels in the cluster are from Spain, but channels
`from Latin American countries are also present.
`The Spanish and Brazilian Portuguese clusters
`appear to be most similar to the entertainment stream
`of the larger English-language YouTube environment:
`video gaming and music content dominate the large
`nodes in this cluster, with interspersed news and
`political content. The Russian cluster 12760,
`in
`contrast, appears closer to the political stream in
`content. While there are large channels producing
`gaming and music content, the frequency of political
`content is higher, buoyed by the constant presence of
`the Russian-language RT.
`The specific language market clusters are not
`totally isolated from English language YouTube, as
`seen during February 2013,
`the month of
`the
`Chelyabinsk meteor. Images of the meteor were
`
`
`
`languages, especially
`
`crossover with non-English
`Polish and German.
`
`4.5 The Politics Stream
`
` A
`
` second major stream of channels is one in which
`the content is primarily political. This stream is
`relatively consistent across the timeline, with four
`contiguous clusters spanning the period from the end
`of cluster 629 to the end of our sample. Cluster 629,
`while containing more news and political content than
`909, still contained elements of popular culture like
`hip-hop music. Cluster 1442 does not share these
`elements, and is defined by its juxtaposition of news
`and conspiracy theory channels. These conspiracy
`theory channels promulgate a wide variety of different
`types of conspiracy theories (e.g. those involving extra-
`terrestrials,
`Illuminati/Freemasons,
`anti-Semitic
`themes, and Christian apocalyptic end-times), though
`they tend toward the conservative end of the political
`spectrum. Channels supporting atheism also appear in
`this stream.
`Clusters 1442, 4099, 8430, and 10737 form a
`continuous chain, with clean breaks between the
`clusters. The