21

This is a follow up to Change Hot Network Questions algorithm


Between 2017-06-06 00:00:02 and 2017-07-20 00:00:10, I have collected 1032 data points, one for each hour on the hour (+ up to 10 seconds), with 2 missing due to internet connectivity issues at 2017-06-07 08:00:02 and 2017-06-28 12:00:02.

Each of these data points is the HTML file for the Stack Exchange homepage. This gives us the sites on the front page of the website - the top 30 HNQ sites.

These HTML files are avaliable here: http://www.filedropper.com/indexfiles

With a simple python script, I've extracted these into a spreadsheet: http://www.filedropper.com/hotquestionsv2

Each row is an hour, and next to the UTC timestamp (column A), one per column, is a site name (columns B to AE). Here's an example row:

+-------------------------+----------+-----------+----------+----------+------------+-------------+---------+---------------+------+--------+--------+-----------+---------+----------+----------+-----+-----+-------+-------+----------+---------------------+---------------------+-----+---------------+-----------+--------+----+-----------+---------------+---------------+
| 2017-06-06 17:00:02 BST | academia | askubuntu | codegolf | codegolf | codereview | electronics | english | graphicdesign | math | movies | movies | parenting | physics | politics | puzzling | rpg | rpg | scifi | scifi | skeptics | softwareengineering | softwareengineering | sqa | stackoverflow | superuser | travel | ux | workplace | worldbuilding | worldbuilding |
+-------------------------+----------+-----------+----------+----------+------------+-------------+---------+---------------+------+--------+--------+-----------+---------+----------+----------+-----+-----+-------+-------+----------+---------------------+---------------------+-----+---------------+-----------+--------+----+-----------+---------------+---------------+

From my data, the top 8 sites are Programming Puzzles and Code Golf, Worldbuilding, The Workplace, Math, Role-playing Games, Science Fiction and Fantasy, English Language and Usage, Academia and Travel.

I selected these and counted the occurrences of each site for each hour (columns AF to AN). Here's an example row:

+-------------------------+     +----------+----------+---------+------+-----+-------+--------+-----------+---------------+
|       Date time         |     | academia | codegolf | english | math | rpg | scifi | travel | workplace | worldbuilding |
+-------------------------+ ... +----------+----------+---------+------+-----+-------+--------+-----------+---------------+
| 2017-07-18 09:00:01 UTC |     |        2 |        6 |       1 |    1 |   1 |     2 |      2 |         2 |             4 |
+-------------------------+     +----------+----------+---------+------+-----+-------+--------+-----------+---------------+

This simply means that at 2017-07-18 09:00:01, there were 2 academia questions, 6 code golf questions etc in the top 30 of the list.

Here's the averages:

+---------+----------+----------+---------+------+------+-------+--------+-----------+---------------+
|  Site   | academia | codegolf | english | math | rpg  | scifi | travel | workplace | worldbuilding |
+---------+----------+----------+---------+------+------+-------+--------+-----------+---------------+
| Average |     1.35 |     4.08 |    1.55 | 1.69 | 1.65 |  1.58 |   0.93 |      1.71 |          1.85 |
+---------+----------+----------+---------+------+------+-------+--------+-----------+---------------+

Clearly PP & CG has way more on the list at any given time, and overall, than the others - every hour I surveyed it had between 1 and 8. The most any other site had at one time was 5.

This is a frustratingly unclear image showing the number of questions in the top 30 of the HNQ list, by site (click for full size):

To simplify it, this is PP & CG vs Worldbuilding:

Here is a stacked area graph:


Around 33% of the HNQ list is taken up by these 9 sites, which make up just 5% of the network's traffic.

13% is taken up by Programming Puzzles and Code Golf, which is only 0.044% of the network traffic.

6% of the HNQ is Worldbuilding, which is only 0.059% of the network traffic.


22 sites make up 81% of the HNQ:

+----------+----------+----------+----------+-------------+-------------+-------+---------+-------+-------+--------+---------+----------+----------+-------+-------+----------+---------------+-----------+-------+--------+-------+-----------+---------------+
|   Site   | academia | aviation | codegolf | cseducators | electronics |  ell  | english | math  | money | movies | physics | politics | puzzling |  rpg  | scifi | security | stackoverflow | superuser |  tex  | travel | unix  | workplace | worldbuilding |
+----------+----------+----------+----------+-------------+-------------+-------+---------+-------+-------+--------+---------+----------+----------+-------+-------+----------+---------------+-----------+-------+--------+-------+-----------+---------------+
| % of HNQ | 4.52%    | 1.95%    | 13.62%   | 2.1%        | 1.98%       | 1.82% | 5.17%   | 5.57% | 1.64% | 2.27%  | 1.8%    | 1.78%    | 1.71%    | 5.51% | 5.28% | 1.42%    | 2.31%         | 1.7%      | 2.24% | 3.12%  | 1.61% | 5.67%     | 6.22%         |
+----------+----------+----------+----------+-------------+-------------+-------+---------+-------+-------+--------+---------+----------+----------+-------+-------+----------+---------------+-----------+-------+--------+-------+-----------+---------------+

Of those 22, Stack Overflow has the worst ratio of traffic to HNQ, with 60% of the traffic and just 2.3% of the HNQ. Code Golf has the worst in the other direction, with 13.6% of the HNQ and just 0.04% of the network traffic.

The best site in the list is Unix & Linux with 1.61% of the traffic and 1.01% of the HNQ.

It appears that weighting on traffic alone is not enough - do we want to do anything about this, and if so, what?


Update 2017-07-30 17:15 UTC

my observations suggest that SE team are possibly running some experiments with [the] HNQ now. In particular, things seem to be different in a how repeated questions from the same site are handled as if the penalty for these has increased (that may be temporarily / experimental)

–– gnat, Jul 27

The red line is where the previous data ended, the green line is the 27th of July. It reached a maximum between those days of 8 questions on the HNQ for 4 hours. For the whole data, up to the end of the 19th of July, the average number of Code Golf questions at any time is 4.09 From the start of the 20th onwards it's 4.45.


Any mistakes in this post, please edit. Any mistakes in the data, comment and I will fix them ASAP.

  • would be interesting to see how many sites (of total 150+) occupy 80% of the list – gnat Jul 20 '17 at 6:51
  • 3
    Looks like we really need to max HNQ amount each site can have at any given time. This will make it more fair for the smaller sites. – ShaWiz Jul 20 '17 at 8:43
  • @ShadowWizard The issue is because PP&CG has so many, whatever number we chose will be the number that appears. 99.5% of the time, there are 2 or more PP&CG questions, 88.1% there are 3 or more, 65.2% there are 4 or more and 36.6% there are 5 or more. If we pick a limit of 2, there will pretty much always be 2 PP&CG questions, so the list would no longer be organic. I'm fine with there being 5 questions at once, but not 37% of the time - HNQ should follow traffic, with favour given to smaller sites (but not much of a bonus). I think the issue is we need to drop the answer boost to a max of 3. – Tim Jul 20 '17 at 12:16
  • 1
    @gnat 22 sites, which make up 76% of the network traffic. See the edit – Tim Jul 20 '17 at 12:18
  • with regards to Stack Overflow it is maybe worth noting that HNQ feature is practically disabled for this site (as explained in more details here): system pushes SO questions off the hot list after several hours. At all other sites things work in a completely different way – gnat Jul 20 '17 at 12:25
  • @gnat it causes the ratio of traffic:HNQ to be 26 for SO. Code Golf is the other direction, it’s 0.003. It’s a little hard to explain but if stack overflow is 26 in one direction, code golf is 340 in the other. I.e. code golf is boosted way more than stack overflow is penalised. – Tim Jul 20 '17 at 12:30
  • I am not sure that comparing these two makes sense because SO is so much different. Frankly I can't understand how to account for SO at all because of the way how differently it is treated by a system compared to all other sites. Another interesting thing I noticed, provided data seems to suggest that about 130 network sites have less than 20 slots in HNQ available to them (there are 100 questions in the list meaning 19% corresponds to 19 slots) – gnat Jul 20 '17 at 12:42
  • I wonder what this would look like if you had data for all 100 questions on the HNQ list. I'd speculate that it would be slightly more diverse, as many low-traffic sites might hit the 50–100 range of the HNQ list somewhat regularly, but only rarely hit the top 30. – Nathaniel Jul 22 '17 at 14:11
  • @Nathaniel I agree, it would be great to have the data. Unfortunately, it was a lot harder to get the data for all 100 on the list. In addition, the top 30 is what visitors to SE.com see, so the lack of diversity there is bad. – Tim Jul 22 '17 at 14:22
  • 1
    Thanks @gnat, I was about to do it myself! – Tim Jul 22 '17 at 14:44
  • @Nathaniel only if slightly. In general I would expect second page to be dominated by very same sites because their questions will occupy it while their hotness score very slowly goes down due to aging factor. The way how things work now is that older questions with grandfathered hotness score block newer ones from getting in – gnat Jul 22 '17 at 18:44
  • ...for example, when I just looked at second page now there were only 11 questions (of 50) with age 12h or less. Overwhelming majority are old farts likely slowly descending from places these took at first page a while ago – gnat Jul 22 '17 at 18:52
  • my observations suggest that SE team possibly is running some experiments with HNQ now. In particular things seem to be different in a way how repeated questions from same site are handled, as if penalty for these has increased (that may be temporarily / experimental) – gnat Jul 27 '17 at 13:16
  • 1
    @gnat my data is still being collected, when I get home I’ll run the chart again – Tim Jul 27 '17 at 15:28
  • 1
    @gnat Bounty started :P I've also edited - on average there were more CG questions in that time (20th onwards)! – Tim Jul 30 '17 at 17:17

You must log in to answer this question.

Browse other questions tagged .