User Generated Content

User Generated Content Video

From the paper “I Tube, You Tube, Everybody Tubes: Analyzing the World’s
Largest User Generated Content Video System”, by Meeyoung Cha¤, Haewoon Kwak†, Pablo Rodriguez¤, Yong-Yeol Ahn†, and Sue Moon:

“…is that caching can be made very efficient since storing only
a small set of objects can produce high hit ratios. That is,
by storing only 10% of long-term popular videos, a cache
can serve 80% of requests.

User Participation

The video popularity and ratings (i.e. the number of view-
ers who evaluated the video) show a strong linear relation-
ship for both UGC and non-UGC, with the correlation co-
efficient of 0.8 for YouTube and 0.87 for Yahoo. This is an
interesting observation, because it indicates that users are
not biased towards rating popular videos more than unpop-
ular ones.

Despite theWeb 2.0 features added in YouTube to encour-
age user participation, the level of active user participation
is very low. While 54% of all videos are rated, the aggre-
gate ratings only account for 0.22% of the total views. Com-
ments, a more active form of participation, account for mere
0.16% of total views. Other Web 2.0 sites also have reported
similar trends on relatively low user involvements [11].

How Content Is Found?

We now examine the Web pages that link to YouTube
videos. Based on Sci trace, 47% of all videos have incom-
ing links from external sites. The aggregate views of these
linked videos account for 90% of the total views, indicat-
ing that popular videos are more likely to be linked. Nev-
ertheless, the total clicks derived from these links account
for only 3% of the total views, indicating that views com-
ing from external links are not very significant. We have
identified that the top five web sites linking to videos in
YouTube Sci are myspace.com, blogspot.com, orkut.com,
Qooqle.jp, and friendster.com; four of them from social
networking sites, and one on video recommendation.

User-generated content, by definition, varies widely in its quality. One may argue that
the natural shape of the popularity distribution of UGC
is truncated (e.g. log-normal), since significant frac-
tion of videos in UGC are of low interest to most users.
For example, UGC is normally produced for small au-
diences (e.g. family members), as opposed to profes-
sionally generated content.

When the number of videos is small, the inefficiencies of the system (due to
filtering effects) are smaller since information can be found
easier.

Search or
recommendation engines typically return or favor a
small number of popular items [15, 36], steering users
away from unpopular ones and creating a truncated
tail. This truncation is more apparent over time since
old non-popular videos are exposed longer to such post-
filtering.

For very young videos (e.g. newer than 1 month),
we observe slight increase in the average requests, which in-
dicates viewers are mildly more interested in new videos,
than the rest. However, this trend is not very pronounced
when we examine the plot of maximum requests. Some old
videos too receive significant requests. In fact, our trace
shows massive 80% of videos requested on a given day are
older than 1 month and this traffic accounts for 72% of to-
tal requests. The plot becomes noisy for age groups older
than 1 year, due to small number of videos. In summary,
if we exclude the very new videos, user’s preference seems
relatively insensitive to video’s age.

While user’s interests is video-age insensitive in a gross
scale, the videos that are requested the most on any given
day seem to be recent ones. Over a one day period, roughly 50% of the top
twenty videos are recent. However, as the time-window in-
creases, the median age shifts towards older videos. This
suggests ephemeral popularity of young videos.

… after a day, 90% of videos are watched
at least once, and 40% are watched over 10 times. After a
longer period of time, more videos gain views, as expected.
One noticeable trend in the graph is the consistent deeps
at certain times (e.g. 1 day, 1 month, 1 year). These
points seem to coincide with the time classification made
by YouTube in their video categorization. From this plot,
we can see that the slope of the graph seems to decay as
time passes. Noting the log-scale in the horizontal axis, this
indicates the probability of a given video to be requested de-
creases sharply over time. … (The date) indicates that if a video did not get enough requests
during its first days, then, it is unlikely that they will get
many requests in the future.

Our results show that second day record gives
an accurate estimation with a relatively high accuracy (cor-
relation coefficient above 0.8). Using the third day record
improves the prediction accuracy, yet, only marginally. Our
results also show a high correlation with the second day
record even for more distant future popularity (e.g. three
months afterwards).

…young videos can change many rank positions very fast,
while old videos have a much smaller rank fluctuation, in-
dicating a more stable ranking classification for old videos.
Still, some of the old videos also increased their ranks dra-
matically. This could indicate that old videos are able to
ramp up the popularity ladder and become popular after a
long time, e.g. due to the Long Tail effects and good rec-
ommendation engines. However, it is hard to conclude this
from Figure 9(a) since a few requests may also result in ma-
jor rank changes.

The gap between the maximum and the top 99 percentile
lines reflects that only a few young videos (e.g. less than 1%)
make large rank changes, indicating that only a very small
percentage of the young videos make it to the top popular
list while the rest have much smaller ranking changes….
. A detailed look at those videos reveals
that those videos did not receive any request during the
trace period, however their ranking was pushed back as other
videos got at least one request. This shows that unpopular
videos that do not receive any request will die in the ranking
chart at a rate of 2000 positions per day.

Better Use of Caching

Caching stores redundant copies of a file near the end user
and has been proven to be extremely effective in many Web
applications. Several factors affect the caching efficiency:
the cache size, the number of users and videos, the correla-
tion of requests, the shifts in popularity, and so on. Here, we
hypothesize a virtual global cache system for YouTube and
assess with real trace how many hits on YouTube servers can
be eliminated. Such cache could be deployed centralized or
fully-distributed.

…we consider the following three conventional caching schemes:
1. A static finite cache, where at day zero the cache is
filled with long-term popular videos. The cache con-
tent is not altered during the trace period.
2. A dynamic infinite cache, where at day zero the cache
is populated with all the videos ever requested before
day zero, and thereafter stores any other videos re-
quested during the trace period.
3. A hybrid finite cache, which works like the static cache,
but with extra space to store the daily most popular
videos.
We populate the static cache with long-term popular videos
accounting for 90% of total traffic. This corresponds to 16%
of Sci videos as in the Pareto Principle.

The results indicate that about 40% of the videos that
are requested daily are different from the long-term popu-
lar videos. However, the corresponding number of requests
toward those videos accounts for only about 20% of the to-
tal requests. In fact, we can see that a simple static cache
that stores the top long-term popular files uses 84% less
space than a dynamic infinite cache solution which stores
all videos, and still manages to save about 75% of the load
in the server. It is worth noting that only about 2% of videos
that are requested every day are newly uploaded ones. We
should also mention that, by storing the most popular daily
requests in addition to the long-term popular videos, a hy-
brid cache improves the cache efficiency by 10%, compared
to the static cache.

Our results indicate that information filtering
is the likely cause for the lower-than-expected popularity of
niche contents, which if leveraged, could increase the total
views by as much as 45%.

Social Media Ads

Promoting user-generated content is promoting your brand. In addition to featuring user-generated content in your social media ads, there are many way you can leverage UGC to improve SEO. For example, your brand can use your social network or buy ads to drive more traffic to positive high-quality user-generated content. It would help raise brand awareness especially if you mention the brand name and website in the message. If you have a budget, spend it on influencer marketing to promote UGC. Hire top social media influencers to help you promote quality user-generated content. It is frequently cheaper and more effective than running your own ads.

The ego-defensive and social functional sources contribute significantly to attitudes formulated about the creation of UGC. The ego-defensive function specifically compels people to protect themselves from internal insecurities and external threats, and the creation of UGC in this sense helps consumers minimize their self-doubts and feel a sense of community. The social function assists consumers in seeking out activities that are perceived as favorable by important others and gives them the opportunity to associate with friends. In relation to the creation of UGC, consumers engage in such actions to connect with others and feel important.