Demystifying Youtube Recommendations — from a user’s point of view
I am a Youtube power user. I spent a lot of time on the platform exploring and consuming interesting, entertaining, educational contents. Every once in a while, I’d come across some amazing gems of discovery 2 or 3 am in the morning. The excitement from these discoveries rival that of my childhood spirit of finding an amazingly shaped rocks in the middle of the forest. I truly truly love the recommender behind the scene that was able to find me these contents.
Now what do people mean when they say “Youtube Recommendations”? If you were like me, I took it that they simply meant an umbrella term, a catch-all if you will, to mean any videos pushed to you by Youtube.
However, there should be more nuance to that. I came to realize this issue after started watching Professor Joseph Konstan’s class on Coursera.
Approaches of Recommenders
There are all different types of recommenders, serving for all different use cases. These recommender approaches can be roughly put into the following categories:
A. Non-personalized and stereotyped recommenders: these are pretty straightforward and simplistic. Think of categories such as hot, trending, most Favorited, most viewed, most liked etc. A good example of these type of recommenders are your typical Twitter trending hashtags, reddit hot, new, controversial posts etc.
B. Product Association recommenders: These are the types of recommenders that would recommend people who liked x also liked y, people who viewed x also viewed y, people who viewed x also bought y. These are again non personal recommenders that can be easily generated through co-occurances of the item in the database. Some prominent examples such as:
C. Content based filters, Collaborative filters and hybrid: These are recommenders are personalized. The gist of the content based recommenders is that they predict what a particular user might like based on what content the user engaged with in the past. The collaborative filters predict what the user like based on what other similar users like.There are many ways to implement these two recommenders. Sometimes they are even combined together to take advantage of the benefits of either. However, on the face of it, it’s actually pretty difficult to know what a given example implemented what type of recommender without looking at their underlying model architecture.
However there are exceptions where the recommenders may call out in general:
Based on this framework, it is already much clearer in understanding what “recommendation” means in greater nuance.
Taxonomy of Recommenders
Professor Konstan’s class also proposed an analytical framework in understanding the use cases around recommenders. Although it is a bit “academic”, it is by far the clearest and the most satisfying framework in thinking about the use cases of recommendation tasks for me. In his framework, he thinks about the recommender feature in the following 7 dimensions:
Domains of recommendation: what is the recommender recommending? Is it recommending content to be consumed (news, music, videos, articles, titles), products to be bought (books, lights, lamps), people to connect (linkedin, tinder, match.com). The content can also take in multiple mode of presentation, it could be one item, a few items or a sequence of items.
Purpose of recommendation: This is to think about what’s the goal of the recommendation. What is it trying to accomplish? Is it trying to build a community (reddit, social network)? Improve sales (Amazon)? Present relevant information or educating the user (News)? These purposes are a bit on the higher level. For specific purpose of a recommender, it may be as detailed as “bring trending news near your current geography that matches with your past 6 months’ topic interest so you can stay informed quickly”.whose opinion is it? Why is it telling me I should do so and so? Is it based on my past browse history or something else?
Consumption Context of recommendation: What is the user doing at the time of consuming the recommendation? Are they sitting around by themselves inside or outside shopping with friends, or waiting idly on the subway train? How does the context of recommendation constrain or add to the recommendation? (think during shopping for your favorite book in Strand book store in NYC while cross checking the listing price on Amazon with limited Wifi coverage). Does the recommender automate, disrupt or require full attention?
Personalization level: Is the recommendation generic to the population, a demographic group or specific to me or people like me? This is a gradient for different use cases. Is it ephemeral: as in relevant to my today’s data, previous year data or is it permanent? Something relevant to me four years ago might not be relevant today anymore.
Privacy and trustworthiness: Is the recommendation trustworthy? Is it telling me something because there’s some biased agenda or political opinion behind? Imagine how manipulative recommenders can be. Think if amazon only pushes to you the products that help it make more money without giving you a list of products that have the highest quality. Maybe it is already doing it?
Recommendation interface: What does the recommendaiton look like? Is it a list? Is it a card? Is it some chat bubble? How does the user interact with the recommendation? Can the user rate, like, dislike, click, share or something else?
Now armed with these two important frameworks, I am ready to critically evaluate the different recommenders available to me as a Youtube power user.
List of Recommenders in Youtube
This is my attempt at listing all the recommenders that I have current access to. I know this won’t be exhaustive since Youtube does a lot of experiments concurrently so it is impossible to know whether the youtube versions is the standard.
My approach here is to primarily focus on the desktop browser accessible version of Youtube. Youtube’s user interface’ structure for content consumption is pretty straightforward. It is primarily consist of some “feed” page and feed item detail page.
The feed page consist of the following:
- Home page feed
- Trending feed
- Subscription feed
- Youtube Originals feed
- Movies and show feed
- Gaming Feed
- Live Feed
The detail page:
- Video detail page
- Playlist detail page
I will go through the home page recommenders and try to answer the question using Konstan’s framework to think through the recommenders on each page. I have to caveat that because of the amount of personalization/experimentation that’s available on the platform. It’s difficult to see what exactly the standard set of recommenders that have been deployed on the platform, that are available to everybody. Also these pages would differ greatly before and after you have logged in.
Why am I not going through all the other pages? Because there are just way too many recommenders on the site.
The non-logged in page recommendation solves a different type of challenge than when the user logged in. For now, the scope of the discussion will mostly focus on recommenders for users who have logged in.
The home page itself is a continuous scroll of all different types of recommender categories. In the following section, I will try to deconstruct these categories.
Recommendations on Home Page Feed
Here are the different recommenders that are present on the home page feed (So far, I saw these 11 types of recommenders):
- 12 recommended videos at the top, show more to reveal additional videos
- Breaking news segment
- Continue watching segment
- Topic specific recommendations
- Channel specific popular uploads
- Recommended channels
- Recently uploaded video recommendations
- Specific channel viewers also watch …
- Youtube Mixes (endless playlists personalized for me)
- From subscriptions, recommended videos
- Live recommendation
12 recommended videos at the top Segment (show more to reveal additional videos, 18 in total)
- Domains of recommendation: Individual videos to be watched
- Purpose of recommendation: If you are a logged in user like me, these set of videos would be the first sequence of videos you’d see. If Youtube recommendation does well, out of all these 12 videos, I’d hold “ctrl” and click open several. These videos should ideally shortcut the user into the “Youtube Rabbit Hole”. All these recommenders are optimized so that users can stay on the Youtube platform as long as possible. the type of data that these recommended videos are based on seem to be mostly historical interested contents as well as contents that other users similar to you have watched. I don’t think anything very diverse would popup in this recommender result (at least nothing I can recall)
- Consumption Context of recommendation: For these types of top of the page recommenders, the user could be anywhere. But for the user to become interested and warped into the rabbit hole, users should ideally by themselves, ready to be entertained, to pass time.
- Personalization level: these videos are highly personalized. They are youtube’s best guesses at what would take you into that rabbit hole.
- Privacy and trustworthiness: So far all these video contents in this section recommended have been very consistent with my perceived past expressed interest level. I trust Youtube’s algorithm in bringing me things that are similar to what I am already interested in.
- The recommendation interface: every recommended item contains these important components: Video thumbnail, video length, video title, channel name, verification badge (if it has one), live badge (if it has one), #views, time it was originally posted. On hovering, video plays through key frames of the video while revealing additional interactions (watch later — appeared twice, save to later, save to playlist, not interested and report)
Breaking news segment
- Domains of recommendation: Individual news story video
- Purpose of recommendation: Gather the latest important news around. “Important” is not very clear as of now since it could be breaking national news or something reaching at a local level. Most of the time this breaking news type of recommender isn’t present on the home page. I often observe this recommender section whenever there’s big fire, political events, national tragedy etc.
- Consumption Context of recommendation: The context doesn’t really change as much for the user. As long as user is bored, Youtube is there for you.
- Personalization level: these videos are not personalized. Seems to me that even if you are just loggedin as a regular Youtuber, you’d get the same national news.
- Privacy and trustworthiness: Most of the news recommended are pretty neutral. They don’t really take sides. I have seen both FOX, CNN coverage news stories in the breaking news section.
- The recommendation interface: Same as all others.
Continue watching segment
- Domains of recommendation: Individual video that you haven’t finished watching
- Purpose of recommendation: Help you pick up where you have left off
- Consumption Context of recommendation: The continue watching segment recommendation only shows up once in a while, though I am pretty sure that I have way more videos I haven’t finished watching than what shows up here.
- Personalization level: They are personalized. These are exactly the videos that you haven’t finished watching.
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
Topic specific recommendations Segment
- Domains of recommendation: Individual video of a specific topic that you’d be interested in
- Purpose of recommendation: You may be of interest of a topic, abstracted from your previous watch history. You can even subscribe to this topic. This topic subscription is a bit ambiguous to me. In addition to individual channels, you can now subscribe to topics.
- Consumption Context of recommendation: Nothing has changed.
- Personalization level: They are personalized. These are videos related to topics of interest.
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
Recommended Channels:
- Domains of recommendation: Channels that you’d be interested in along with newest videos from that channel
- Purpose of recommendation: Increase your channel subscription count
- Consumption Context of recommendation: Nothing has changed.
- Personalization level: They are personalized. These are channels that best fit your
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
Channel specific popular uploads Segment
- Domains of recommendation: Popular videos from specific channel where you have shown intense interest
- Purpose of recommendation: Popular videos are usually pretty good at getting engagement
- Consumption Context of recommendation: Nothing has changed.
- Personalization level: Not personalized
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
Recently uploaded video recommendations Segment
- Domains of recommendation: These are videos that match with your interest but are uploaded within the last 5 days. These are videos that may or may not come from your subscription feed
- Purpose of recommendation: Give newly uploaded videos a window for exposure so new contents can be spread out
- Consumption Context of recommendation: Nothing has changed.
- Personalization level: They are personalized. These are videos that match with your interest
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
“Specific channel viewers also watch” segment …
- Domains of recommendation: These are videos from channels you may or may not have subscribed to but are watched by viewers from the channels that you are interested in.
- Purpose of recommendation: These are videos that are similar specifically to the channel of your intense interest. It kind of works like Amazon’s buyers of an item also bought type of recommender. It will be fairly accurate for users to discover related videos.
- Consumption Context of recommendation: This segment appears once in a while. It doesn’t show up for all channels that you visit. It only shows up when you are intensely interested in the videos of a channel
- Personalization level: They are personalized.
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Same as all others.
Youtube Mixes (endless playlists personalized for me)
- Domains of recommendation: Music videos that are automatically stitched together into youtube playlists.
- Purpose of recommendation: Autoplay Youtube music videos that can keep you on the platform
- Consumption Context of recommendation: These mixes can be played out loud for parties or ambience music or listened to privately at work/study. However, if you want to listen to these music on the go outside similar to how you’d consume Spotify, maybe less so.
- Personalization level: They are personalized. Playlist is personalized, though the recommendation itself doesn’t seem to work very well, not as well as other videos.
- Privacy and trustworthiness: Not applicable
- The recommendation interface: These playlist/mixes are a bit different. They show the depth of these mixes, usually in the upwards of 50+ items.
From subscriptions, recommended videos
- Domains of recommendation: These are individual videos from all your subscriptions. It would surface up recommended videos from your subscribed channels with also a recency filter on top.
- Purpose of recommendation: I have subscribed to 175 channels on Youtube. Assuming these channel update .5 video per week on average. That’s 87.5 new videos per week. They would accumulate so fast that would eventually end up making channel subscription obsolete, since there’d be too much videos to watch from just your subscriptions and not all of them are that interesting. Recommender is here to specifically help that use case where your subscription is useful and it can help you filter out videos that best match with your interest.
- Consumption Context of recommendation: Catching up with the newer video that come from your subscriptions
- Personalization level: They are personalized but limited to videos that you have subscribed to
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Nothing has changed. Same as above.
Live Recommendations
- Domains of recommendation: Livestreams from different channels
- Purpose of recommendation: Livestream is a different category than regular videos. These recommenders will share view exposures from video views.There are far less live streams than there are videos on youtube. These are also effective ways to have multiple people engaged on the video content
- Consumption Context of recommendation: The livestream seems to be mostly on streaming of events and music or let’s play content.
- Personalization level: They are personalized. Topics seem to be smilier to what you have watched before
- Privacy and trustworthiness: Not applicable
- The recommendation interface: Similar to video recommendation interface with the exception that live badges and concurrent watchers are also reported.
Conclusion
First of all, thank you for bearing with me to actually get to this section of the article. As you can see that there are a lot of different type of recommenders that each fulfilling a different mission/use case than the others. From closely examining these in production recommenders, we can gather roughly the following:
- Recent/fresh contents are important and are given ample attention
- Youtube recommenders can abstract from your watch history topics that you’d be interested in. The topics are fairly diverse and different from each other
- Popular videos are also given enough real estate on the Youtube home feed
At last, hopefully by this point, we’d know in detail that when we say Youtube recommendation, it isn’t as blackbox as we’d originally imagined. It is in fact capable of being unbundled and deconstructed. We can see that the concept of “youtube recommendation system” is in fact made up of many many different individual recommender units.
