> 99% algorithms are public already,

What?

You need to back up that statement, because I do not understand how it's at all true.

Or do you somehow have access to Facebook's news feed algorithm, or the Youtube recommendation algorithm?

Maybe they mean 99% of known algorithms. Instead of 99% commercial visibility of a select few??

The fundamental math is almost always SVD (or related), some type of clustering, or some "fancier" minimization.

These companies aren't inventing new math, they're just interpreting/biasing the results differently.

For example, an ML algorithm wouldn't be the secret sauce. Most will result in the same answer. The training data, and its weights, are the secret sauce. Would this consider the training data as part of the algorithm? I would assume not.

I see you keep editing, so I'll respond to the latest version of your comment:

Assuming an ML-based methodology (and I would not personally assume that), yes, the training data, the resultant weights, and the algorithm itself would be of interest, though at least the last two would be necessary.

Sorry, I reworded a bit. I didn't realize anyone would respond so quickly!

LOL, this is ridiculous.

All computing is "fundamental math".

That there are some known set of algorithms in the mathematical toolbox that are used is irrelevant.

The question is, how are those tools put together, what are their inputs, and how are those inputs weighted.

algorithm: a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer.

Is user data, that's used to train, part of the process? I don't think so. It's an input to the algorithm. In the case of ML, the input is the only thing that matters.

Semantics, but I think it's incredibly important. It's the difference between posturing and actually affecting change.

Frankly, at this point you're just playing semantic games.

The goal: To understand why the system is recommending what it's recommending.

"Algorithm" is a short-hand for that goal.

Everything else is just clever wordplay to try and deflect the conversation.

> "Algorithm" is a short-hand for that goal.

I think it's the incorrect word. The "algorithm" requires a few pages of code. "Understanding why it's recommending" requires you get, and sift through, terabytes of user data from Youtube. They're very different.

That's an interesting claim that's based on nothing but your own supposition.

How about we get Youtube to show us their algorithm and associated weights (assuming an ML-based approach, which is a huge assumption on your part) and then we can find out if that's true.

> assuming an ML-based approach, which is a huge assumption on your part

What more reasonable alternative do you see, with the existence of my personal recommendation feed? I'm not aware of any personal recommendation systems that don't use ML. Do you have any examples in mind?

I believe the point GP was making was that the "algorithm" is not the important part, and is actually reasonably easy to reverse engineer if it's not just copied wholesale from an existing algorithm. The important part is the weights used in the algorithm. Since the weights are input, and usually semi-unique to the user, they would be considered private information.

For example, I could tell you that the "sorting algorithm" I use to foo my bars is ksort[0]. But that's not really meaningful information without also knowing what the keys are.

[0]: https://www.php.net/manual/en/function.ksort.php

> I believe the point GP was making was that the "algorithm" is not the important part

Not the only important part.

> Since the weights are input, and usually semi-unique to the user, they would be considered private information.

Cool, then as an individual user, let me see that, too.

why couldn't they be anonymized?