A Reverse-engineering YouTube Recommenders?

WWGD

Science Advisor
Gold Member
4,203
1,759
Hi,
This is just out of curiosity. I have been hearing music from YouTube for a while. It seems no matter
where I start I end up in the same two songs, as sort of "Attractors".
Can I use the "path" generated by a specific song as a seed ( first one I listen to in a given internet session) and use the somg-path to reverse-engineer the underlying song recommender? EDIT: I am not sure how the LEvels issue works, so I will take any suggestion/idea.
 

scottdave

Science Advisor
Homework Helper
Insights Author
Gold Member
1,577
589
My guess is there are 2 (or more) types of algorithms at work. One could be based on your own listening history and try to recommend songs similar to ones which you like. There would need to be some method to decide what is similar and what is not similar. Another could be based on an association - what other people listened to after listening to a particular song. Or perhaps a method similar to Google's pagerank system, to figure which ones have the strongest connections.
 
663
294
The following links are to some open-access works co-authored by John T. Riedl, who was an IEEE Fellow, a University of Minnesota Professor, and a great man who had a special knack and compassionate motivation for imagining how people could co-contribute meaningfully in collaborative efforts. I think his sensibilities in that regard are part and parcel of what led him to be among the pioneers of recommender systems.

https://spectrum.ieee.org/computing/software/deconstructing-recommender-systems
https://www.researchgate.net/publication/215470714_Evaluating_collaborative_filtering_recommender_systems
http://www.ra.ethz.ch/cdstore/www10/papers/pdf/p519.pdf
http://files.grouplens.org/papers/ec00.pdf
https://apps.dtic.mil/dtic/tr/fulltext/u2/a439541.pdf

John was a computing buddy of mine when we were in school (early-to-mid '80s), and he once had me co-present some ideas, mostly his own, for machine processing of natural (human) language.

The professors mostly thought the project ideas were too ambitious (possibly partly because we were both undergrads), but they let us get started on it, looking at ways to recognize grammatical constructs such as parts of speech, how to construct context-based semantic disambiguators based partly on analysis of proximities and frequencies of co-lexemes, syntactical facts, etc..

The faculty members questioned what uses we might find for such systems -- I said that minimally, we could enhance spelling checkers to recognize that "to, too, and two" don't always add up to six (that example of course works better in speech than in writing), or with a bit more effort, we could highlight possibly incorrect usage of "affect" versus "effect". John's response was more general, along the lines of "I don't really have much in mind yet about what the applications might be; I just think it'd be interesting; I'm curious about how far we could get with it just as a general purpose AI tool project."

John had previously asked me to write a command line interface to invoke a full-screen editor he'd written in PL/1, modeled after one he'd previously written in COBOL. In response, I'd written a few hundred lines of CLIST code to faithfully mimic the invocatory syntax and file allocation consequences of the IBM TSO EDIT command (a line-oriented, non-full-screen editor), making his editor more conveniently accessible to everyone.

John not only recognized and appreciated the abilities of others; if you were in his vicinity he would find some way for you to pitch on something he was working on, or he'd contribute helpfully to your work -- https://www.richrelevance.com/blog/2013/08/07/remembering-john-riedl/
 

WWGD

Science Advisor
Gold Member
4,203
1,759
My guess is there are 2 (or more) types of algorithms at work. One could be based on your own listening history and try to recommend songs similar to ones which you like. There would need to be some method to decide what is similar and what is not similar. Another could be based on an association - what other people listened to after listening to a particular song. Or perhaps a method similar to Google's pagerank system, to figure which ones have the strongest connections.
I have heard them called collaborative filtering : when your "similars" ( people known to like songs you like too) and content-based filtering : recommendations based on songs one has listened too. Weird thing is that the same seed of two initial songs lead me through different paths/songs.
 

Want to reply to this thread?

"Reverse-engineering YouTube Recommenders?" You must log in or register to reply here.

Physics Forums Values

We Value Quality
• Topics based on mainstream science
• Proper English grammar and spelling
We Value Civility
• Positive and compassionate attitudes
• Patience while debating
We Value Productivity
• Disciplined to remain on-topic
• Recognition of own weaknesses
• Solo and co-op problem solving
Top