Popularity Growth Patterns of YouTube Videos
A Category-based Study
Shaiful Alam Chowdhury and Dwight Makaroff
Department of Computer Science, University of Saskatchewan, Saskatoon, SK, S7N 5C9, Canada
Keywords:
Workload Characterization, Multimedia Applications, Content Distribution, Time-series Clustering.
Abstract:
Understanding the growth pattern of content popularity has become a subject of immense interest to Internet
service providers, content makers and on-line advertisers. This understanding is important for the sustainable
deployment of content distribution systems. A significant amount of research has been done in analyzing the
popularity growth patterns of YouTube videos. Unfortunately, little work has been done that investigates the
popularity patterns of YouTube videos based on video object category. In this paper, we perform an in-depth
analysis of the popularity pattern of YouTube videos, considering video categories. We find that the time
varying popularity of different YouTube categories are different from each other. For some categories, views
at early ages can be used to predict future popularity, whereas for some other categories, predicting future
popularity is a challenging task and requires more sophisticated techniques (e.g. time-series clustering). The
outcomes of these analyses can be instrumental towards designing a reliable workload generator, which can be
further used to evaluate different caching policies and distribution mechanism for YouTube and similar sites.
1 INTRODUCTION
YouTube and other user generated content (UGC)
sites have altered the way people watch video on the
Internet. YouTube was the 4
th
most accessed Inter-
net site in 2007 (Cheng et al., 2007), and its use was
increasing over time in a power-law manner. Re-
cent studies continue to support two central observa-
tions: 1) increasing number of videos and users (Ding
et al., 2011; Siersdorfer et al., 2010) and 2) dissatisfy-
ing experiences of users in watching YouTube videos
(Khemmarat et al., 2011). Other recent studies (Gem-
ber et al., 2011; Labovitz et al., 2010; Maier et al.,
2010) suggest that YouTube is the most bandwidth in-
tensive service of today’s Internet, and it accounts for
20-35% of Internet traffic.
Much research has been done investigating re-
quest characteristics from both client (Gill et al.,
2007; Zink et al., 2009) and server perspectives
(Borghol et al., 2011; Cha et al., 2009; Ding et al.,
2011; Figueiredo et al., 2011) in order to enable im-
proved service. However, none of this earlier work
considered the types of video objects. This aggregate
data may not tell the whole story.
A proper understanding of YouTube’s workload
will aid in the design of new systems, as well as ca-
pacity planning, and network management for similar
types of systems. The methodology we have devel-
oped is useful for UGC sites that have a single cache
for the region of requests captured.
In this paper, the time-varying global viewing pat-
terns of a sample of YouTube videos from their up-
load time are analyzed, considering video category.
1
We present the results of one data collection period
(5 months of views of videos uploaded in 2 consecu-
tive days); a previous dataset showed similar charac-
teristics and is not evaluated here. Our results show
that different categories exhibit different viewing pat-
terns in terms of overall popularity and detailed pop-
ularity over time. We confirmed that the number of
views of the popular videos follows a Zipf distribu-
tion for most categories, whereas views of the unpop-
ular videos follow a heavy tail distribution. We also
show that time-series clustering can be successfully
used to understand the growth patterns for the cate-
gories where early popularity cannot be used to pre-
dict future popularity of a video.
These observations contribute to a better under-
standing of the popularity dynamics of YouTube
videos, enabling realistic testing scenarios for devel-
oping and evaluating various design parameters for
UGC sites. While the request patterns for different
categories may vary around the world, our dataset and
1
As defined by the uploader
233
Chowdhury S. and Makaroff D..
Popularity Growth Patterns of YouTube Videos - A Category-based Study.
DOI: 10.5220/0004372802330242
In Proceedings of the 9th International Conference on Web Information Systems and Technologies (WEBIST-2013), pages 233-242
ISBN: 978-989-8565-54-9
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
analysis provide a case study that shows that cate-
gory differences persist in global access patterns, and
therefore will exist in each region. Our analysis en-
ables the development of category-specific workload
generators which can be combined to form the input
for simulators and prototype systems. While devel-
oping and evaluating a comprehensive workload gen-
erator remains as future work, we have a strategy for
generating synthetic requests on a category basis and
present preliminary results which match reasonably
well for two categories: News and Music.
The remainder of the paper is organized as fol-
lows. Related work is described in Section 2. Section
3 explains the data collection methods of our study.
We discuss the characterization of the request patterns
in Section 4, and use the information from views over
time to develop a workload generator for two cate-
gories in Section 5. Section 6 provides conclusions
and future work.
2 RELATED WORK
Previous request characterization and video popular-
ity analysis has been used to investigate the feasibil-
ity of different content delivery streaming techniques,
and to design and evaluate caching policies/systems
for UGC sites. Our work leverages the best practices
in the previous literature to investigate category pop-
ularity over time.
YouTube video request traffic was captured at the
packet level at the University of Calgary over a 4
month period (Gill et al., 2007). They investigated
video popularity properties, usage patterns, and trans-
fer behaviours as measured from the client edge of the
distribution network. The traces examined contained
data from both completed and incomplete requests.
Their analysis suggests that appropriate caching deci-
sions not only can improve end user experience, but
also reduce network bandwidth usage.
Another study (Zink et al., 2009) observed the
traffic of YouTube videos between a university cam-
pus and the YouTube server. Approximately 25%
of the videos in the trace were requested more than
once, leaving a long tail in the distribution. Three
different content delivery techniques were analyzed:
P2P based distribution, proxy caching and local
caching. Proxy-caching outperformed the other tech-
niques, and P2P based distribution sometimes exhib-
ited worse performance than local caching.
These two results can be biased by the measure-
ment locations which appropriately restrict the con-
text of the studies and the solutions that are pro-
posed. For instance, it is claimed that video requests
in YouTube follow a Zipf distribution (Gill et al.,
2007), which is different from other works that con-
sider global request patterns. For our purposes, global
access patterns are essential.
2.5 million YouTube videos were obtained using
related video links (Cheng et al., 2007). Access pat-
terns of the popular videos did follow a Zipf-like dis-
tribution, in spite of having a heavy-tailed section
in the distribution curve. Data collected indicated
that the YouTube network is similar to small world
networks, and P2P techniques could be successfully
applied, contradicting earlier findings (Zink et al.,
2009). Their dataset is likely to be biased to popular
videos because of the crawling approach, and popu-
larity over time is not investigated in detail.
A recent approach to investigate growth patterns
in YouTube video requests was to use Google charts
to collect views over time (Figueiredo et al., 2011).
They analyzed the time-varying viewing patterns of
popular videos, deleted videos and randomly selected
videos. Popular videos usually experience a huge
number of views on a single peak day or week. Unfor-
tunately, using the Google charts API is not sufficient
to have a proper, fine-grained understanding of the dy-
namics of video popularity as Google charts API al-
ways returns 100 data points, regardless of video age.
Recent work was done on nearly 30,000 videos,
collected by using the recently uploaded standard feed
provided by the YouTube API (Borghol et al., 2011).
Their collection procedure claims to have an unbiased
dataset; the Most Recent standard feed returns video
information randomly that are uploaded recently.
Most of these videos experienced their peak pop-
ularity within fewer than six weeks of their uploading
time. Video collection based on keyword search is
shown to be biased to popular videos. This observa-
tion suggests that in order to accurately characterize
the viewing patterns of YouTube videos, the method
of data collection is important.
3 DATA COLLECTION
No prior work measures the daily views of different
categories of YouTube videos from the first day of
their uploading time. We modified previous unbiased
data collection methods (Borghol et al., 2011) since
we speculate that the first week since uploading de-
serves more investigation, even though this may ex-
pose day-of-week effects. Moreover, similar numbers
of videos from all the categories are needed for appro-
priate comparison between different categories. Mul-
tiple crawlers were deployed to obtain data used in
our analysis. Since the crawler obtained information
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
234
from the API, the crawler location is irrelevant.
(1) Most Recent Crawlers. 15 different crawlers
were deployed on March 3
rd
, 2012 (a Saturday), col-
lecting video IDs for 15 different categories,
2
by re-
stricting the Most Recent queries to a specific category
for each crawler. All crawlers collected video infor-
mation for 24 hours, ensuring that subsequent video
views began on the first day of their lifetimes. The
Most Recent standard feed provides video informa-
tion randomly, reducing bias to particular classes of
videos. A similar procedure was followed on March
4
th
, 2012. After two days, a total of 71,208 videos’
information was obtained. Depending on the server
load, the YouTube API returns at most 100 videos’
information for each request every one or two hours,
limiting the dataset size.
(2) Video View Collection Crawlers. Video view
collection using two separate crawlers was started
from March 4
th
, 2012 and March 5
th
, 2012. This
continued for 149 consecutive days (approximately
5 months). The crawlers ensured a 24-hour differ-
ence between view collections. Normalization was
performed on the first day’s views. Due to network
connection failures, some video views on days 20 and
58 of the measurement period were not captured. For-
tunately, those days are not that important for most of
the videos, as most of the significant events occur at
the very early age of a video. After normalization,
147 day’s views are analyzed.
After 149 days, the number of videos in the dataset
fell from 71,208 to 47,711 (an average deletion rate
of 33%). Manually sampling of the data set revealed
that a large percentage of the deleted videos had copy-
right infringement issues. Table 1 shows the summary
of our dataset. Howto, Film, Entertainment and Tech
videos experience the highest deletion rates. Analysis
of deletion rates is left as future work though deletion
rates for all categories decrease over time.
(3) Uploading Rate Crawlers. Another crawler
was developed that collected category names of
videos provided by YouTube’s Most Recent stan-
dard feed. The crawler ran for 5 months, starting
from February 2
nd
, 2012 and collected approximately
365,000 unique videos’ information. This allows us
to estimate the short-term current category-specific
uploading rates. While not an accurate representation
of the entirety of YouTube, it does give some insight.
2
http://support.google.com/youtube/bin/
answer.py?hl=en&answer=94328
Table 1: Categories and Number of videos.
Category Number Number Deleted
of videos of videos videos
(Day 1) (Day 149) Pct
Howto 4773 1772 62.87
Film 4654 2346 49.59
Ent. 4991 2528 49.34
Tech 4942 2682 45.73
Games 4711 2966 37.04
People 4310 2730 36.65
Autos 4714 3245 31.16
Comedy 4744 3467 26.91
News 4623 3432 25.76
Travel 4918 3698 24.80
Sports 4812 3733 22.42
Music 4774 3477 21.93
Nonprofit 4624 3691 20.17
Education 4710 3801 19.29
Animals 4908 4143 15.58
Total 71208 47711 33.00
4 VIDEO REQUEST ANALYSIS
4.1 Time-varying Category Popularity
Figure 1 shows the cumulative distribution functions
(CDF) of time-to-peak for the videos from different
categories with at least 100 views; a video with a very
small number of views might contribute unfairly to
the understanding of the actual growth pattern of a
category. One consequence of this restriction is that
the number of videos in each category is significantly
reduced, down to 42% for News and Sports and 18%
for Animals and Travel. We define time-to-peak as the
day in which a video experienced the most views as in
previous work (Borghol et al., 2011). Time to reach
peak popularity is not the same for all categories.
News and Sports categories follow a similar distri-
bution and the time to reach peak popularity for these
two categories is the shortest. Approximately 85% of
News and Sports videos reach peak popularity within
the first 4-5 days of their lifetimes. As well, in every
category, between 50% and 60% of the videos expe-
rience their peak viewing on Day 1. Other categories
such as Music, Film, Howto, Tech and Education fol-
low similar patterns and many videos in these cate-
gories reach peak popularity much later.
The other categories follow similar distributions,
and peak distributions of these categories lie within
the previous two groups. The significance of time-
to-peak can be enhanced by Figure 2 which depicts
PopularityGrowthPatternsofYouTubeVideos-ACategory-basedStudy
235
Figure 1: CDF of time-to-peak.
Figure 2: Percent of total views over time.
the CDF of percent of total views over time for all
the videos in a subset of categories. Music and Film
videos experience relatively fewer views early in their
lifetime. Film videos follow an almost constant view-
ing rate for the entire measurement period. News and
Sports videos, however, experience a significant por-
tion of the total views early.
It is important to understand if the peak day differs
significantly from other days of a video’s lifetime in
order to determine if our previous statistic is helpful.
Figure 3 shows the complementary cumulative distri-
bution function (CCDF) of the most distant day x after
the peak such that the views on day x is at least 50%
of the peak views, defined as follows:
x = max(i) : view(i) ≥ 50% ×view(peak) & i > peak
(1)
Figure 3: CCDF of time-after-peak.
where view(i) is the views on day i and view(peak)
is the number of views on the peak day. Only videos
with more than 100 views are considered. Figure 3
shows the peak day as a unique point in the lifetime
of videos for faster-growing categories (e.g., News
and Sports). These categories experience a popular-
ity burst, and quickly decline to a lower viewing rate.
Many Music, Film, Howto, Education and Tech
videos that reach peak popularity comparatively lately
do not have that drop in their popularity (Figures 1
and 3), so time to reach peak popularity is propor-
tional to the active lifespan of a video. For example,
over 75% of the News and Sports videos never ex-
perience half of their peak days’ views after the peak
day (Figure 3), but fewer than 50% for Film and Tech
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
236
videos have this characteristic. The stability of Film
and Tech videos suggests that a longer measurement
period would increase the difference between these
categories and News/Sports.
We are also interested to know if the categories
that reach peak popularity faster than others also ex-
perience differing numbers of views. Figure 4 depicts
the 95
th
percentile of views of all categories over time.
We show the 95
th
percentile to remove the potential
effect of outliers. This shows which categories have
a minimum percentage of popular videos (5%) during
the first 100 days of the data collection and the relative
popularity of the categories for those popular videos.
The last 49 days of the collection period are virtually
identical to days 50-99 in terms of this measure.
These graphs illustrate how viewing patterns
of different categories change throughout the early
part of their lifetimes. Although the most similar
dataset collected (Borghol et al., 2011) shows that
the views of Music category exceeds all other cate-
gories within their 8-month measurement period,
3
our
dataset shows that popular News, and Sports videos
enjoy higher viewing rates than any other types of
videos for the first couple of days since publication.
Figure 4 suggests that almost all categories have at
least 5% of their videos that experience a high ini-
tial viewing rate; the difference is that after these few
peak days, views for most of the categories become
very low, except Music and to a lesser extent, Film
and Tech videos. The results indicate the variations in
active life spans of different categories.
Although similar results can be observed from
the average views per day (Figure 5), this can be
misleading because of the high variance of views.
The higher early average views of Sports videos than
News videos is due to the most popular video in the
entire dataset, which happens to be a single enor-
mously popular Sports video with almost 24 times
that of the second most popular Sports video.
4.2 Fractions of Popular Videos
The percent of videos with different views of the
YouTube categories are shown in Table 2. Only ap-
proximately 10% of the Music videos enjoy fewer
than 10 views; this value is over 30% for Howto,
People, Autos, Comedy, and Travel. Music, News,
Sports, and Film contain most of the popular videos
in our dataset (> 1.11% with over 10,000 views). The
most unpopular videos are in the Travel category, fol-
lowed by Comedy and Animals. Only 0.44% of the
People videos had more than 10,000 views, in spite
3
We collected category names of the videos which had
not been deleted by running another crawler.
Figure 4: 95
th
Percentile of views per day.
Figure 5: Time varying average added views.
of the highest uploading rate (shown later). Although
uploaders currently upload more UGC videos, users
are still not attracted to UGC videos compared to
UCC (user copied content) videos.
4.3 Current Uploading Rate
In order to design a request generator for YouTube,
it is important to know the category uploading rate.
In 2007, Music was in the top position in number
of uploaded videos followed by Entertainment, Com-
edy, Sports and Film (Cheng et al., 2007). Manual
sampling revealed that these categories are dominated
by UCC rather than UGC, so most of the videos in
YouTube were actually UCC.
Figure 6 shows the current uploading trend of
YouTube videos obtained by crawler 3. We see that
the uploading trend in YouTube has changed over
time. The People category is at the top position with
approximately 24% of all the new videos, which was
at the 6
th
position in 2007, only 8% of all the videos.
Samples from the People category contain compara-
tively more UGC objects than other categories.
PopularityGrowthPatternsofYouTubeVideos-ACategory-basedStudy
237
Table 2: Percent of popular videos.
Category ≤10 views 11 to 100 101 to 1000 1001 to 10000 10001 to 100000 > 100000
Pct Num Pct Num Pct Num Pct Num Pct Num Pct Num
Music 10.44 363 48.72 1694 32.87 1143 6.38 222 1.29 45 0.29 10
News 18.85 647 39.57 1358 31.61 1085 8.42 289 1.4 48 0.15 5
Sports 20.79 776 46.0 1717 26.12 975 5.97 223 1.04 39 0.08 3
Tech 22.56 605 47.28 1268 24.61 660 4.85 130 0.63 17 0.07 2
Film 23.06 541 49.53 1162 20.84 489 5.46 128 1.07 25 0.04 1
Entertainment 27.77 702 46.88 1185 20.61 521 3.88 98 0.75 19 0.12 3
Howto 43.79 776 34.59 613 17.04 302 4.01 71 0.45 8 0.11 2
Nonprofit 24.11 890 48.04 1773 23.49 867 3.85 142 0.46 17 0.05 2
Education 24.73 940 48.83 1856 21.7 825 4.34 165 0.37 14 0.03 1
Animals 25.59 1060 56.48 2340 15.52 643 2.05 85 0.34 14 0.02 1
Games 27.51 816 49.36 1464 19.08 566 3.44 102 0.51 15 0.1 3
People 29.52 806 49.93 1363 17.69 483 2.42 66 0.4 11 0.04 1
Autos 30.57 992 41.45 1345 23.17 752 4.07 132 0.68 22 0.06 2
Comedy 32.33 1121 51.08 1771 14.08 488 2.08 72 0.35 12 0.09 3
Travel 33.75 1248 48.89 1808 15.44 571 1.76 65 0.14 5 0.03 1
Figure 6: Category Uploading Rate (365,000 videos).
4.4 Category Popularity Distributions
Figure 7 shows the Rank-frequency distribution for
the 6 categories that showed the most interesting pat-
terns. Previous studies (Abhari and Soraya, 2010;
Cheng et al., 2007) showed that although requests for
popular YouTube videos follow a Zipf-like distribu-
tion, a Weibull distribution fits better because of the
heavy tail section, which indicates a large number of
very unpopular videos in YouTube. After considering
video categories, only News videos follow a Weibull
distribution for the first 80% of the videos, because of
the comparatively flatter head section of News access
pattern. This is consistent with fetch-at-most-once be-
haviour (Gummadi et al., 2003), as would be expected
in watching news videos. For all other categories, re-
quest distributions of popular videos follow Zipf dis-
tributions and the heavy tail sections of the categories
can be fit with a Weibull cutoff, as can be seen with
the high goodness of fit statistic (R
2
). The number of
videos that exhibit Zipf behaviour differs between the
categories, showing different-sized tails.
Another measure that we calculated was the
CCDF of total views over the measurement period.
There were a substantial number of videos in certain
categories that had at most 1 view. This can skew the
popularity measures. The HowTo and Autos category
had 17% and 12.6% of videos with at most 1 view,
respectively, while 9% of HowTo videos had 0 views.
There is a section of completely unpopular videos that
get published, but never viewed. Figure 8 shows the
CCDF of the total views for a selected number of cat-
egories. We truncate the x-axis to see the behaviour
of views for unpopular videos more clearly. Enter-
tainment is used as an example of a group of cate-
gories that had very similar CCDFs (Entertainment,
Games, People, Education and Tech). The shape of
the distribution of total views is very similar in these
categories, but that of views over time is not. Music
has very few videos below 20 views, but HowTo has
almost 50% of the videos below 20 views.
5 TOWARDS A WORKLOAD
GENERATOR
5.1 Predicting Popularity
As an approach to predict future popularity of videos,
Pearson’s correlation coefficient (Equation 2) is cal-
culated between the added views
4
at different snap-
shots of the measurement period.
4
Added views is the number of views on a particular day
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
238
(a) News (b) Music (c) Entertainment
(d) Film (e) People (f) Comedy
Figure 7: Number of views against rank for categories.
Figure 8: Selected CCDF of total views.
r
xy
=
n
∑
x
i
y
i
− (
∑
x
i
)(
∑
y
i
)
q
n
∑
x
2
i
− (
∑
x
i
)
2
q
n
∑
y
2
i
− (
∑
y
i
)
2
(2)
A high correlation coefficient between early views
and and the rest of the period implies that prediction
of future views of individual videos is achievable (Sz-
abo and Huberman, 2010). We got very encouraging
results for some of the categories including Sports,
Travel, Howto, Tech and Games.
5
However, for other categories like Film, News,
Entertainment the coefficients are very poor, indi-
cating the significant changes in the set of popular
videos. Music shows a bit different characteristics
though, if we take first 10 days as our first snapshot.
5
Sports is 0.99 for the first day’s views and the rest of
the measurement period
5.2 Time-series Clustering
This category variation led us to model the growth
patterns differently. Three-phase characterization
(Borghol et al., 2011), does not work for the cate-
gory specific modeling, as the number of videos that
are at or before their peak phases in a particular day
are very different between first few days and last few
days in our measurement period. We thus decided to
investigate whether the growth patterns of videos in
a specific category follow similar shapes. This ap-
proach can be considered as a time-series clustering
problem and becomes a challenging problem as dif-
ferent videos reach peak popularity at different times.
Inspired by a study on viral videos (Broxton et al.,
2010), we translate all the time-series so that the x-
axis is centred on the peak day, since most of the sig-
nificant events happen around the peak periods.
Another challenging issue is to select the appro-
priate time-series clustering algorithm. We are par-
ticularly interested to identify similar shapes of the
views per day, regardless of the time to peak. More-
over, the algorithm should not be affected much by
outliers. We selected K-SC clustering (Yang and
Leskovec, 2011), which has been found to be accurate
in identifying the growth patterns of other Web con-
tent. Unlike K-means clustering, K-SC cluster cen-
troids are not distorted by outliers. Instead of consid-
ering Euclidean distance between the curves, K-SC
applies a scale and shift invariant distance metric (Chu
and Wong, 1999). We evaluated the performance of
K-SC algorithm for only two categories: Music and
PopularityGrowthPatternsofYouTubeVideos-ACategory-basedStudy
239
News. The clustering was performed only for the top
2000 videos in order to present more accurate results.
Figure 9 shows the six clusters for Music videos
found by K-SC. Forcing K-SC to select fewer than six
clusters drops the accuracy significantly, as we lose
some of the interesting patterns. However, more than
six clusters does not significantly improve the accu-
racy as we observe the repetition of similar clusters.
Figure 9: Music-clusters.
The cluster shapes for News videos (Figure 10)
are very similar to Music (except very little differ-
ence between cluster (a) and (e) in Figures 9 and
10 respectively).
6
However, the numbers of videos
in each cluster differ between these two categories,
complementing our earlier findings. 46% of Music
videos are contained within the slower-decaying clus-
ters; this drops to 15% for News videos.
An important question that must be answered is
whether a particular cluster is more biased to popular
videos than others. This can be answered by taking
the average of the rank values of all the videos in a
cluster. The central limit theorem suggests that the
average rank of each cluster of videos should be 1000
if it is not popularity-biased. For News videos, the
average rank values are very similar for each cluster
(near 1000). For Music videos, the clusters with com-
paratively slower decay contain more popular videos,
with average rank values of approximately 700. Pop-
ular Music videos observed a sharp decay with less
frequency than popular News videos.
5.3 Performance of K-SC
In order to evaluate the performance of K-SC, we de-
signed a synthetic workload generator for News and
6
Matching clusters are not in the same position in the
graphs
Figure 10: News-clusters.
Music videos. The synthetic data should show simi-
lar characteristics to the empirical YouTube data if the
clustering of K-SC is accurate.
The workload generator can be described as fol-
lows. A rank value is assigned to each of the 2000
videos as suggested by the chosen distributions for
Music and News respectively. Then centroid/cluster
is assigned to the videos based on the distribution we
observed earlier. We also imposed a little bias for
the popular videos before selecting the appropriate
cluster in order to match our observed average rank
value. Although for News videos the time-to-peak
distributions are very similar for each of the clusters,
we found very different results for Music videos, as
shown in Figure 11. We consider these peak distribu-
tions separately in our request generator.
Figure 11: Peak distribution for music videos.
We test similarity between the synthetic and em-
pirical data from four different perspectives: 1) The
total view distribution, 2) time-to-peak distribution, 3)
Average daily views over time, and 4) 95
th
percentile
of views over time.
Figures 12 and 13 indicate very good matches
between synthetic and empirical data for metrics 1
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
240
Figure 12: Time-to-peak distributions.
Figure 13: Popularity distributions.
and 2, which does not in itself indicate high ac-
curacy of K-SC. We imposed the distributions for
these two cases from our observations, i.e., fixed
peak and Zipf/weibull distributions. Metrics 3 and
4 show, however, that the clusters found by K-SC
algorithm for both categories represent most of the
videos growth patterns (Figure 14 and 15, respec-
tively). Similar daily average views on a particular
day indicates that view distribution among videos on
that particular day are similar both in the empirical
and synthetic data. These results show a smaller num-
ber of outliers in both of the categories.
6 CONCLUSIONS AND FUTURE
WORK
In this paper, we analyzed global daily viewing pat-
terns of a representative subset of YouTube videos
from their time of publication until they were 5
months old. We discovered significant time-varying
popularity differences between categories.
Most videos exhibit their peak viewing day very
Figure 14: Daily views over time.
Figure 15: 95th percentile of views over time.
soon after publication and then there is a decay; rel-
atively few videos ever return to near their peak pop-
ularity. We determined that video categories which
reached their peaks later were more stable. This is ex-
pected and matches our intuitions. We developed an
analysis method that permits quantification of these
differences on a particular dataset. The confirmation
of Zipf distributions for the total views of popular
videos in nearly every category indicates that caching
would be effective.
We were also able to determine the relative trends
of viewing patterns of videos within categories over
the first few months of their lifetimes. Some cate-
gories contain a non-trivial number of videos which
are still popular 5 months after upload date, whereas
other categories have viewing patterns which dwin-
dle to nothing. Some categories have videos which
exhibit stationary behaviour that allows prediction of
which videos will remain relatively popular based on
their early views. Popularity changes around peak
time can be captured by appropriate time-series clus-
tering. While we use a dataset from YouTube, issues
regarding the scale and deployment make direct ap-
plicability to YouTube impractical. Multiple regional
PopularityGrowthPatternsofYouTubeVideos-ACategory-basedStudy
241
caches are needed to satisfy the demand and regional
differences (Brodersen et al., 2012). Our methodol-
ogy and analysis could be used to help design, con-
figure, and deploy any category specific UGC site.
As future work, we are in the process of building a
complete workload generator that encompasses more
aspects of user-generated content video requests. In
particular, we will incorporate category-specific in-
troduction of new content over time to drive simula-
tions and/or prototype content distribution networks
to evaluate different design policies for storing and
delivering videos.
REFERENCES
Abhari, A. and Soraya, M. (2010). Workload Generation
for YouTube. Multimedia Tools and Applications,
46(1):91–118.
Borghol, Y., Mitra, S., Ardon, S., Carlsson, N., Eager, D.,
and Mahanti, A. (2011). Characterizing and Mod-
elling Popularity of User-Generated Videos. Perfor-
mance Evaluation, 68:1037–1055.
Brodersen, A., Scellato, S., and Wattenhofer, M. (2012).
YouTube Around the World: Geographic Popularity
of Videos. In World-Wide Web 2012, pages 241–250,
Lyon, France.
Broxton, T., Interian, Y., Vaver, J., and Wattenhofer, M.
(2010). Catching a viral video. In IEEE Data Min-
ing Workshops, pages 296–304, Sydney, Australia.
Cha, M., Kwok, H., Rodriguez, P., Ahn, Y., and Moon, S.
(2009). Analyzing the Video Popularity Characteris-
tics of Large-Scale User Generated Content Systems.
IEEE/ACM Trans. Netw., 17(5):1357–1370.
Cheng, X., Dale, C., and Liu, J. (2007). Understanding
the Characteristics of Internet Short Video Sharing:
YouTube as a Case Study. Technical report, Cornell
University, arXiv e-prints.
Chu, K. K. W. and Wong, M. H. (1999). Fast time-series
searching with scaling and shifting. In ACM PODS
1999, pages 237–248, Philadelphia, PA.
Ding, Y., Du, Y., Hu, Y., Liu, Z., Wang, L., Ross, K., and
Ghose, A. (2011). Broadcast Yourself: Understanding
YouTube Uploaders. In ACM IMC 2011, pages 361–
370, Berlin, Germany.
Figueiredo, F., Benevenuto, F., and Almeida, J. (2011). The
Tube over Time: Characterizing Popularity Growth of
Youtube Videos. In ACM WSDM 2011, pages 745–
754, Hong Kong, China.
Gember, A., Anand, A., and Akella, A. (2011). A Compar-
ative Study of Handheld and Non-handheld Traffic in
Campus Wi-Fi Networks. In PAM 2011, pages 173–
183, Atlanta, GA.
Gill, P., Arlitt, M., Li, Z., and Mahanti, A. (2007). Youtube
Traffic Characterization: A View From the Edge. In
ACM IMC 2007, pages 15–28, San Diego, CA.
Gummadi, K. P., Dunn, R. J., Saroiu, S., Gribble, S. D.,
Levy, H. M., and Zahorjan, J. (2003). Measure-
ment, modeling, and analysis of a peer-to-peer file-
sharing workload. In ACM SOSP 2003, pages 314–
329, Bolton Landing, NY.
Khemmarat, S., Zhou, R., Gao, L., and Zink, M. (2011).
Watching User Generated Videos with Prefetching. In
ACM MMSYS 2011, pages 187–198, San Jose, CA.
Labovitz, C., Iekel-Johnson, S., McPherson, D., Oberheide,
J., and Jahanian, F. (2010). Internet Inter-Domain
Traffic. In ACM SIGCOMM 2010, pages 75–86, New
Delhi, India.
Maier, G., Schneider, F., and Feldmann, A. (2010). A First
Look at Mobile Hand-held Device Traffic. In PAM
2010, pages 161–170, Zurich, Switzerland.
Siersdorfer, S., Chelaru, S., Nejdl, W., and Pedro, J. S.
(2010). How Useful are Your Comments?: Analyz-
ing and Predicting YouTube Comments and Comment
Ratings. In World-Wide Web 2010, pages 891–900,
Raleigh, NC.
Szabo, G. and Huberman, B. (2010). Predicting the popu-
larity of online content. CACM, 53(8):80–88.
Yang, J. and Leskovec, J. (2011). Patterns of temporal vari-
ation in online media. In ACM WSDM 2011, pages
177–186, Hong Kong, China.
Zink, M., Suh, K., Gu, Y., and Kurose, J. (2009). Char-
acteristics of YouTube Network Traffic at a Campus
Network - Measurements, Models, and Implications.
Computer Networks, 53(4):501–514.
WEBIST2013-9thInternationalConferenceonWebInformationSystemsandTechnologies
242
