I suggest you ...

Make "sort by rating" better

I have two ideas that I would like to share to improve search relevance.

Have you ever noticed how low-view count videos tend to have either much higher ratings or much lower ratings? But once a video passes 1 million views, it always hovers about 80%?

That is because counting statistics matches the Poisson distribution. If 300 people click "thumbs up," then the "uncertainty" on 300 is sqrt(300)= 17. So the estimated number of people who liked it from those who watched it is 67% likely to be within 317 to 283.

My guess is you already incorporate view count into "relevancy" searches, but I think that you should incorporate the number of thumbs up into the % people who liked a video for "top rated" searches. I.e., I would like the % people displayed be the LOWER estimate, which better matches intuition. The formula for this is as follows:

Let u = the number of people who thumbs up the video
Let d = the number of people who thumbs down the video
Then, this "lower estimate" on the % people who liked the video is:

(u/(u+d))*(1 - sqrt( (1/u) + (1/(u+d)) ))

This value incorporates the Poisson error from each count to produce an estimate for the lower bound on the % people who liked it. I think this is a much cleaner result, useful for searches. For example, if:

u = 2890
d = 669

Then, the lower-bound % is given by:


Whereas the true % is given by 81%.

If instead, we had:

u = 10
d = 1

Now, this is very uncertain! This is reflected in the lower bound estimate:


Versus the center value of:

10/(10+1) = 91%

Do you see? Now, if I currently click to sort by "top rated", what I see is a bunch of videos with a few thumbs up, all hitting 100%. But if instead, you guys were to incorporate Poisson error, using my formula above (which, e.g., turns 81%->79% for high count, and 91%->51% for low count), then you would be giving a much better (and statistically rigorous) estimate for the true % of people who like a video (well... the lower bound on that %, which is important).

Now, of course, you wouldn't have to display a "lower bound" estimate next to each video, but at least incorporate that into search results, so I don't see videos with only 3 thumbs up when I search!


Similar to above, but more straightforward. At each view count range, look at the % of people who like a video. E.g., 0-1000 views, 1000-10000 views, etc.

Then, when I click "sort by top rated", the videos should be listed according to their PERCENTILE within their particular view-count range. E.g., at 1 million+ views, say the 99th percentile (videos having ratings better than 99% of all other 1million+ view videos) videos are shown first. But, if a video with 10 million+ views has a lower rating than this video, but happens to be a better PERCENTILE within its own view category (say 99.9th percentile), then that video is shown before.

This is simply because among videos with equal ratings, if one video has much more views than the other, then obviously that video should appear first in search! So this percentile method is an easy way to show that - by incorporating the fact that % people who like a video tends to go down as the views increase.


- Anon

289 votes
Sign in
or sign in with
  • facebook
  • google
    Password icon
    Signed in as (Sign out)
    You have left! (?) (thinking…)
    Anon Ymous shared this idea  ·   ·  Admin →


    Sign in
    or sign in with
    • facebook
    • google
      Password icon
      Signed in as (Sign out)
      • BKFL commented  · 

        This is Pornhub not math class

      • Molly Montreux commented  · 

        Jan. 9, 2018, 8:04 PM EST

        I read it. Some day, I might even understand it.
        A little.

        A slightly bigger issue, it seems, is whether the people who
        run pornhub could find the time or energy to do anything,
        about anything, even if they gave a shit, and I'm not about to
        bet the house on _that_.

      • RoseLes commented  · 

        LOL! Nobody read this shit.

      Feedback and Knowledge Base