Every Google product I've used has had hilariously awful recommendations.

For a company that has so much of my data and indexes the entire internet, it sure is surprising how bad they are at recommendations.

> What's the point of a recommendation system if it doesn't even recommend new video of the same category I like?

Because YouTube doesn't care what you want--YouTube cares what it can monetize.

I suspect that YouTube gets very little money from showing you classical performances but gets lots from showing you the latest "Funny Cat Video".

You are the product, not the customer.

Once you internalize that, stuff that Google does makes sense.

> Because YouTube doesn't care what you want--YouTube cares what it can monetize.

> You are the product, not the customer.

> Once you internalize that, stuff that Google does makes sense.

They published their algorithm. They rank items according to predicted watch-time. The other poster is probably correct, in that many people use these long classical music videos as background music over and over (i.e. high watch time) and so they're highly ranked.

Interesting. When and where did "they publish[] their algorithm?"

RecSys 2016

Covington, Paul, Jay Adams, and Emre Sargin. "Deep neural networks for youtube recommendations." Proceedings of the 10th ACM conference on recommender systems. 2016.

Nvidia furthermore implemented/released it for TensorFlow, and for their new Recsys engine:

https://ngc.nvidia.com/catalog/resources/nvidia:wideanddeep_...

The thing is, it doesn’t even show me funny cat video anymore. It’s the same classical music video recommended over and over again.

To be honest, the more I think about it, the more I don’t know what money is YouTube trying to earn.

It’s probably just a dumb algorithm.

I think it has a grasp of revisits, I forget the exact terminology but if you rewatch stuff I tries to predict will you like to watch it again. Some stuff is not worthing multiple times though, once is enough. But music videos can indeed be watched multiple times.

I think I heard somewhere that a lot of kid videos rank in the top stratosphere of view counts on YouTube because kids just play stuff on repeat. There's probably a lot of repeat plays going on across YouTube.

It's because a large segment of classical music watchers are using it as a background music playlist, and listen to the same recordings repeatedly.

They should be able to distinguish between users who always play the same video in a loop vs. users who never watch anything twice (or anything intermediate between the two extremes) and adjust the recommendations based on that.

But maybe their recommendation engine is hyper-focused on the matrix-completion framework and doesn't use any information other than which users watched which videos.

The recommendation system is not "dumb" in the opinion of its designers. Perhaps it is performing exactly as intended.

In the same way that online advertising needs to display millions of ads to reach only hundreds of people, the recommendation system may not be intended to work at 100% efficiency.

YouTube "recommendations" work as far as I can tell. If you take two people searching for some subject area on YouTube, and only one of them knows how to search past the initial "popular" recommendations, each person will produce very different results.

If neither person is a skilled searcher (most people are not), they both end up with the same results. Those videos receive thus more and more views. Videos with millions of views help Google sell online ad services. If views were distributed more evenly across all videos, would that make advertisers more interested or less interested.

What's the point of a recommendation system that does not recommend a new video in the same category. To promote certain videos to attract large audiences and thus draw more interest from potential advertisers. More advertisers bidding on fewer popular videos would seem to favour higher prices paid to Google.

Some of the best videos I have seen on YouTube have less than 500 views, some even in the single digits. I found them through automated searches, not "recommendations".

Google makes information accessible with the belief that relevance is directly proportional to popularity. The company was founded on this idea.

This isn’t how digital ad inventory works. Advertisers don’t bid to win a slot that all viewers of a particular video will see. They bid to win a view. A million views on one video, or the same million impressions split across a thousand videos doesn’t change the supply/demand equation like you suggest. Buying an ad on the most popular video isn’t like buying an ad on the Super Bowl, because not all viewers are seeing the same ads.

A video with only a few hundred views is not going to produce much YouTube ad revenue. In order to sell the idea of making money from YouTube ads, one needs to convince ad buyers and video uploaders that videos can receive millions of views.

Same problem with music, I thought I somehow broke their engine because it keeps feeding me some random hit songs I liked 5 years ago and does it for the past 5 years. You would think they want to show me something new for their engagement metrics but I guess not.

Contrary data point: I watch a lot of classical music on YouTube and it gives me consistently excellent recommendations.

(Although I only take recommendations from the front page, and ignore the playlists entirely.)