Question

Trying to implement a similar approach towards our view counts for our web app.

Reading this article: http://mashable.com/2012/06/25/why-do-youtube-videos-freeze-at-301-views/

And watching: http://www.youtube.com/watch?v=oIkhgagvrjI

What approaches is YouTube taking to determine whether a view is valid? For example, not coming from bots, views services, or a user trying refresh the page several times. I know they probably have several approaches towards this. But looking to get started.

Answer 1

This is slightly difficult to answer and I am not from youtube. But I can take a stab at a few things to help you think about this.

When should you declare a view? There are several options:

1. The moment the user clicks on a video link and data starts flowing. [Simplest engineering solution but not really a valid measure]
2. If the user has watched the first 25% (or 30% or enter your number here) of the video. This could also be changed to say the user watched 40% of the video with scrubbing.
3. If the user has watched the entire video through to the end. Too conservative. Someone may stop watching at the last 5 seconds because say credits are rolling.
4. The user has watched the entire pre-roll advert (perhaps an advertisement client is only interested in this!) before going on to watch the video

There are also aspects of whether the video view is human/automated?

1. Are you getting too many views from the same location at a rate that is not humanly possible?
2. Are your video views showing a very unlikely pattern [say all views stop at exactly 45 seconds or at 50% of the video or all views are always to the end] even if the rate of arrival is not very fast. A human will have variations in the viewing pattern.
3. What are the sources that link to your views. Are you getting views from different sources?
4. Some rules on what was the previous video view, what is the next video view can also add to the detection of bots. [Say videos are being watched in alphabetical order or order that they were presented in a search. you know most likely this is an automated program just going through a list. ]

Then you can combine rules with location, OS, browser, device etc trying to stream. It gets more complicated than that after the initial set of rules. But I think you will get the gist of it.