ties, YouTube videos keep attracting
views throughout their lifetimes. The
rate videos attract views may naturally
differ among videos, with the less-pop-ular likely marking a slower pace over a
longer time.
These two notably different user-popularity patterns are a consequence
of how users react to content on the two
portals. On Digg, articles quickly become obsolete, since they often link to
breaking news, fleeting Internet fads,
or technology-related themes with a
naturally limited time for user appeal.
However, videos on You Tube are mostly
found through search, since, with the
sheer number of videos constantly being uploaded, it is not possible to match
Digg’s way of giving each promoted story general exposure on a front page. The
quicker initial rise of video view counts
can be explained through the videos’ exposure in YouTube’s “recently added”
section, but after leaving it, the only way
to find them is through keyword search
or when displayed as related videos next
to another video being watched.
The short fad-like popularity life
cycle of Digg stories (a day or less) suggests that if overall user activity on Digg
depends on time of day, a story’s popularity may grow more slowly when fewer
visitors are on the site and increase
more quickly at peak periods. For YouTube, this effect is less relevant, since
video views are spread over more time,
as in Figure 1. Figure 2 outlines the
hourly rates of user digging, story submitting, and upcoming Digg story promotions as a function of time for one
week, beginning August 6, 2007. The
difference in rates may be as much as
threefold; weekends showed less activity, and weekdays appeared to involve
about 50% more activity than weekends.
It was also reasonable to assume that
besides daily and weekly cycles, such activity also involved seasonal variations.
Moreover, in 2007, Digg users were
mostly located in the UTC- 5 to UTC- 8
time zones (the Western hemisphere).
Depending on the time of day a submission was made to the portal, stories
differed greatly in the number of initial
diggs they received. As we expected,
stories submitted during less-active
periods of the day accrued fewer digs
in the first few hours than stories submitted during peak hours. This was a
natural consequence of suppressed
figure 3. correlation of digg counts on the 17,097 promoted stories in the data set older
than 30 days. a k-means clustering separates 89% of the stories into an upper cluster; the
other stories are a lighter shade of blue. The bold line indicates a linear fit with slope 1 on
the upper cluster, with a prefactor of 5. 92 (Pearson correlation coefficient of 0.90).
104
Popularity after 30 digg days
103
102
101
101
102
103
Popularity after one digg hour
figure 4. Popularity of videos on the 30th day after upload vs. popularity after seven days.
The bold line with gradient 1 is fit to the data.
105
Popularity after 30 days
104
103
102
101
100
100
101
102
103
104
105
Popularity after seven days
digging activity at night but might have
initially penalized interesting stories
that were otherwise likely to be popular. For instance, an average story promoted at 12 p.m. received approximately 400 diggs in the first two hours and
only 200 diggs if promoted at 12 a.m.
That is, based on observations made
after only a few hours after a story was
promoted, a portal could misinterpret
the story’s relative interestingness if it
did not correct for the variation in daily
user-activity cycles.
Since digging activity varies by time,
we introduce the notion of digg time
measured not in seconds but in num-
ber of diggs users cast on promoted sto-
ries. We count diggs only on promoted
stories because this section of the por-
tal was our focus, and most diggs (72%)
were to promoted stories anyway. The
average number of diggs arriving at
promoted stories during any hour day
or night was 5,478 when calculated
over the full six-month data-collection
period; we define one digg-hour as the
time it takes for so many new diggs to
be cast. As discussed earlier, the time
for this many diggs to arrive took about
three times longer at night than during
the day. This “detrending” allowed us
to ignore the dependence of submis-
sion popularity on the time of day it
was submitted. Thus, when we refer to
