Compile | AI technology base (rgznai100)

Participate in | reason_W


It’s easy to see how popular video websites are, so there are some popular vloggers, such as Papi Jiang, who is very popular with the business leader, and papitube’s Po owners.

So, how do these Internet celebrities make their videos popular quickly? What standards do the websites recommend videos to the masses? What algorithms are used?


In the case that video websites do not disclose their algorithms, how can Internet celebrities who upload videos grasp the routine in the process of video distribution and steadily produce high-quality video content for a long time? Is there a path that actually works? Or really can only rely on three vulgar content to constantly challenge the lower limit of the vast audience?


Welcome to the first part of our parsing of the YouTube algorithm. Let’s start with how an animated video producer at YouTube fought YouTube’s algorithm for a long time. Let’s see how, without knowing anything about YouTube’s algorithm, he worked backwards from several months of experience in operation and promotion to deduce the factors that influence YouTube’s algorithm: Viewing duration, traffic volume, access speed, access duration, start session, upload frequency, duration session, end session, etc.


Now, let’s follow Matt Gielen to explore the six secrets behind YouTube’s recommended videos.




Body:

Whether it’s a feature film, a stage play, a TV show, or whatever video is currently streaming online, as long as you’re creating content for some form of distribution agency, its distribution mechanism can make or break a lot of your work.


For example, if you’re working on a TV show, you expect it to be successful. You’d better know when it’s best to break for commercials, how to promote them more effectively, which channel fits your content better, how many viewers that channel has, and so on.


However, if you’re Posting videos on YouTube, it’s harder, because the most valuable point about its distribution mechanism is how YouTube’s algorithm works. After all, everything about algorithms is hard to understand. What’s more, YouTube doesn’t even disclose which variables its algorithm takes into account.


But even with very limited data available, we want to look inside this giant black box to figure out how it works. Some data is very important to the algorithm, and having access to it (thumbnails and title impressions, user visit history and behavior, viewing duration, etc.) can greatly improve the transparency of the algorithm. But unfortunately, we don’t have access to it.


But we did what we could with the data we could get our hands on. My former colleague Jeremy Rosen and I spent more than six months looking at data from the channels Frederator owned and operated, trying to figure out as much as we could about YouTube’s algorithm.

There are a few things I should mention before we begin. In this article, we will collectively refer to many of YouTube’s promotional algorithms (Recommended, Suggested, Related, Search, MetaScore, etc.) as “YouTube algorithms”. There are many differences, but one thing they all share is that the optimization goal is “viewing time” (= number of views × average viewing time).



For more information on “viewing duration”, see another article by the author

WTF Is Watch Time?! Or How I Learned To Stop Worrying And Love The YouTube Algorithm

As we all know, the success of a video depends on how long viewers watch it. In this article, I’ll detail my observations on some of the core variables that affect “viewing time.”

Watch the time


First of all, “viewing time” does not mean the number of minutes watched. As we discussed earlier, viewing duration is a combination of the following items:

  • Traffic and speed of access

  • The access time

  • Start the conversation

  • Upload the frequency

  • Duration of conversation

  • End of the session

Basically, each of these items is related to how well your channel and video performs: how often viewers come back (to start a page-visit session), and how long they stay.

In order for your channel and video to accumulate any variable values in the algorithm, you first need to get views. In order for these videos to be “successful” (i.e., to reach 50% or more of the audience within the first 30 days), you need to get a lot of views in the first minutes, hours, and days, which we call access speed.

Next, let’s look at two variables in action.


Traffic and speed of access



When analyzing Frederator’s “access speed”, we found that the average cumulative views of videos grew exponentially as the percentage of subscribers visited in the first 48 hours increased:

       

Based on this finding, we dug a little deeper: When we used this “access speed” rule to predict whether a video would perform well, we were 92 percent accurate.

In fact, the average cumulative view of a video is more correlated with the percentage of users who subscribed in the first 72 hours.



These graphs and correlations make it abundantly clear that “visits” and “speed of access” have a direct and significant impact on the overall success of videos and channels.

Moreover, there is evidence that this effect is also apparent in reverse: low “access speed” negatively affects not only the current video, but also the previous video and the following video.

The graph below shows that if Frederator’s last upload had a low “access speed” in the first 48 hours (low is defined as less than 5% of subscribers accessing the video), the next upload will also be negatively affected.

       

The data confirmed a Matthew Patrick in the video (https://www.youtube.com/watch?v=HLJQ0gFHM8s) referred to in the theory. His theory suggests that if one of your videos isn’t viewed by a large number of subscribers, YouTube won’t give your next upload a large subscriber recommendation weight. Or because you didn’t do a good job of up-front uploading, channel access is low, which in turn causes distribution mechanisms to distribute your content to a smaller audience. But whatever the cause, the sad result is the same.

Evidence suggests that another important effect of low “access speed” on new uploads is that it also hurts your video library’s overall traffic.


In the first chart below, the blue line is the number of visits from subscribers in the first 48 hours, and the red line is the seven-day rolling average percentage of visits from subscribers in the first 48 hours compared to the channel’s overall traffic. The second chart shows the number of video visits that day as a percentage of the channel’s overall visits.



Both charts show one thing: When the percentage of subscribers accessing your newly uploaded videos and/or channel library videos drops, so does the number of channel visits overall.

That is, with this algorithm, YouTube actively promotes channels that appeal to the channel’s core audience, while actively penalizing channels that don’t.



The access time



We found that the next largest variable that had a significant impact on the algorithm was the “access duration.” Access duration refers to the length of time viewers stay on a single video page.

This variable has a lot of weight. In our data, you can see a clear turning point: on The Frederator Channel’s data this year, videos that are more than eight minutes long on average were visited 350 percent more in the first 30 days than videos that are less than five minutes long.


The graph below shows the relationship between the average cumulative views of videos on Frederator’s channel and the average duration of those videos.




Note: We don’t consider video data that lasts longer than 8 minutes (because that would be more than 8 minutes once the video is watched).

We also found that the longer the visit, the better the video performed.

The graph below shows the average number of visits in the first seven days for videos of less than five minutes (1), five to ten minutes (5) and more than 10 minutes (10) :

       

The diagram below also means the same thing, but has been extended from 7 days to the entire life cycle.

       

In addition to these findings, we also have a less conclusive conclusion that extending video time improves data access performance.

Frederator has a kidzone channel that posts three to four videos a week (3, 10, 30 and 70 minutes). We noticed that even for some older videos uploaded to the channel library, the first 48 hours of a 70-minute video were viewed much more than other videos. In addition, the 70-minute video is the same as the average length of any other video on the channel.

We recommend that they only post 70 minutes of video per week. Since we adopted our strategy, the average number of visits to the Kidland channel has increased by 500,000 per day, while the number of video uploads has decreased by 75% in the last six weeks. It’s a surprise, I know.


Starts a session, lasts a session, and ends a session


This research Is largely based on my previous published research, WTF Is Watch Time? !


WTF Is Watch Time?! Or How I Learned To Stop Worrying And Love The YouTube Algorithm


Please refer to my above research for details. I will not repeat it here, but simply review the three concepts.


Start the conversation
This is the number of people who go to your YouTube page and start a session, starting with one of your videos.


This shows why the first 72 hours of visits from your subscribers are so important. Subscribers are people who can watch your videos on day one. They are also most likely to click on a thumbnail of that channel because they are familiar with your brand.

Last session
Time is how long the user watches your video and then stays on the page after watching it.


End of the session
This is how often you leave YouTube and end a page session while watching or after watching one of your videos.




Algorithm theory




YouTube’s algorithm focuses on how well channels are promoted, not how well individual videos are promoted.

YouTube’s algorithm shows what they want from the channel:


  • It keeps people coming back to the platform

  • It keeps people on the platform for a long time

The following three graphs give evidence for this theory.


The first chart shows the percentage of visits by 48-hour subscribers and the number of visits by individual videos in 7 days. It tells us that if a lot of users start a Youtube page session with your video (that is, launch session), your video will get a lot of traffic. If subscribers reach a certain threshold, the number of visits becomes exponential:


   

The second chart shows the average number of daily visits relative to the percentage of the channel’s five-day rolling subscribers.

    

This means that if you can consistently get a large number of user launch sessions (a five-day rolling average), the algorithm will increase the number of daily views of your videos sent to the channel’s entire video library.

The last chart shows the average daily visits as a percentage of subscribers and the channel’s five-day rolling visits as a percentage.


     

This suggests that there is a correlation between channel persistence and the number of visits, which are in turn expressed as the percentage of subscribers who visit.

So, let’s say you have a game channel with 100,000 subscribers, upload 6 videos per day, and each video gets 5% of subscribers. Then your rolling average will be a steady and modest 5%. That means you’ll get about 30% of your subscribers per day (or 30,000 per day or 600,000 per month). If we assume that you have 1 million subscribers, those numbers turn out to be 300,000 views per day and 6 million views per month.


We don’t think the math is deceptive. This means that YouTube is choosing channels to promote based on some deterministic metric, and then as the algorithm promotes that channel, it gets a corresponding number of visits.


But note that this analysis is purely theoretical!



Scoring algorithm


Here we will reverse crack these algorithms and reconstruct them. With 15 variables and the best estimate of their weight, we create a scoring algorithm.

Here are the variables we used:


                                       

These variables are used to develop the algorithmic factors of the scoring algorithm.

The graphs below show the actual effects of these factors.

     

     

The graph below shows this in more detail.

   

In case you’re wondering, here’s a (very) rough estimate of the weight of the algorithm’s variables:

     

     

     

However, without more data, we cannot yet determine what type of regression to use in calculating correlations.



Analysis of YouTube’s (current) algorithm



According to our data, there are at least six secrets:

  1. YouTube uses algorithms to determine how many views each video and channel gets.

  2. Successful channels focus on a specific type of content/idea.

  3. Once a channel has established a successful content genre, it should not explore too much.

  4. High-priced content creators will never succeed on YouTube, so this segment will never fully embrace YouTube.

  5. Personalised shows/channels will always be the dominant type of content on the platform because they are the “specific type of content” people want to watch.

  6. New channels that can’t be channeled from outside YouTube will have a hard time driving traffic.

In summary, we believe that the algorithm is designed to promote channels that are able to gather and maintain a fan base of viewers.


If you want to be successful on YouTube, our best advice is that you should focus on a niche market and make videos as long as possible on a single topic of 10 minutes or more.

Mind you, these are my personal notes. YouTube gets a lot of criticism for its algorithms, but I hope they don’t see this as a negative post.


Through this research, MY understanding of YouTube’s algorithmic engineers has deepened. After all, they’re trying to deal with more than a billion people around the world with varying interests every month. When you stop and look at all of this in perspective, it’s amazing how clever Youtube’s algorithms are, designed to achieve its business goals while preventing abuse and keeping the platform healthy.




Matt Gielen is the former vp of programming and audience development at Frederator Networks. Matt managed the team that built the Frederator Networks channel, the largest animation network in the world. In addition, he led the team that created and programmed the Frederator Networks channel on YouTube. This article is their experience through data research and analysis.