Artists: the Pandora algorithm

[NADRIP]

I'd like to start a post on this topic.  I have a general understanding of how this algorithm works and, in general, I believe it's intelligent to try to use computer aided decision making, so support the attempt to make helpful algorithms.

That said, algorithms tend to be heavily Pareto Principle for their success, that being that it's pretty easy to capture 80% of songs, or 80% about any particular song, with just 20% effort.  The process of developing and inputting the algorithm is the effort and for most songs that process will yield much of the important information about the song.

But still just talking theoretically, you really should be running multiple algorithms that take very different approaches to understanding songs.  Ideally, you'd find an algorithm that did an excellent job of capturing the 20% that your main algorithm doesn't do well on, and, vice-versa, whose weak areas were strong areas for your main algorithm.  This is similar to the financial strategy of hedging, so that at all ties with a perfectly hedged algorithm, one of the 2 algorithms will always generate value in its answer, and many times they both will.  Of course, it's very hard to find another algorithm which acts as a perfect hedge.  But in that case, by using multiple algorithms all with completely different approaches to understanding songs, then you would still get a result similar to the financial strategy of diversification.

Ok, now talking specifically, I notice a couple of limiting aspects of your algorithms is that, first, the beginning of your input process is to assign songs to genre-specific experts, based on a very antiquated way of thinking about songs, that a song is, say, (mainstream) country or (mainstream) rock or blues or dream pop or some variants like punk rock.  And then for each of these segments there is an expert on that genre who is going to input the characteristics of the song according to the norms of that genre according to their personal subjectivity.  So, if a song were, say, somewhere between mainstream rock and punk rock then depending on luck it could get two very different assessors with two very different concepts of norms and two very different subjectivities.  That's bad enough just for a single song, but then compounding this further, the way the Pandora algorithm works is that it wants to categorize future songs based on how prior songs by that artist have been categorized.  If one were to look at a financial statement, one would undoubtably see a disclaimer that historical results are no indication of future results or similar, but your algorithm seems to take the opposite view, becoming less flexible as it gathers historical information and wanting to jam all future songs into a genre which may not have even been the best choice to start with.

Now let's look at some realities of American music where your algorithm with the above characteristics won't do well.  Songs that I come out with or collaborate on primarily use an "Americana" style. However, Americana is primarily lryic-focused, ie the lyrics are more poetic than typical lyrics in many song genres but Americana allows these lyrics to be paired with a wide variety of musical styles, basically any style that is rooted in America, so while the lyrical genre of Americana songs remains Americana, the musical style can drift anywhere from, say, singer-songwriter/folk/country to rock to R&B to blues to pop, and even to genres like religious music. I also put out both male and female fronted songs. So most likely your internal team has people pegged to specific musical styles not grouped as Americana (lyric-focused) performers, so if I say Bob Dylan or Woody Guthrie they'll probably say folk and if I say Willie Nelson and the rest of the outlaws they'll say country and if I say John Mellencamp or Tom Petty or Roy Orbison or the Grateful Dead they'll try pegging the song to different styles of rock, or if I say Lana Del Rey they'll peg to dream pop.

But actually a lot of Americana performers have songs which are consistently Americana in a lyrical sense but musically don't stay within your algorithm's genre boundaries.  And for such singers, some of their listeners mainly prefer Americana lyrics so will be flexible as to musical genres, just as the performers are, and for listeners that prefer to stay rutted in the older understanding of songs, then they may like some songs by a particular Americana performer that happen to fall within their preferred musical genre and not like other songs that don't.  In this tradition, currently I have songs out or imminently forthcoming with musical styles range from different types of country, folk, singer-songwriter, indie pop, dream pop, and alt-rock, and even a polka song, and will certainly explore other styles rooted in America in the future.  And your algorithm approach is definitely going to struggle with that.  But this breaking of genre lines as your algorithm defines them has been happening since at least the 1950s, like when Dylan went electric, so it's not like I started it.  There are still people writing to past concepts of genres, but those concepts are and should be evolving.

Anyway, I did use me as an example but I'm really most interested in having a productive discussion about the algorithm Pandora is using now, it's strengths and weaknesses, and what, if anything, could be done to improve the current approach (because there's always room for improvement).  I welcome your thoughts.

mod edit: changed title for clarity

[ChrisPandora]

Hey @NADRIP 👋

We're happy to hear that your music has been included in the Music Genome Project.

I understand your concerns. Regarding genre and artist associations, the music analysts don't draw those analogies. When we analyze a song, we qualify the musicological components of the song. Based on this data, the matching algorithm will create a station populated with artists with similar data.

If you're unfamiliar with our analysis process, here's an article outlining how we do it.

Please note that this information is not able to be manually changed.

You should also know that your station will most likely get better over time, as more listeners give feedback to the songs that play on your station.

I hope this helps to shine a little light on what we do to get your music ready for radio 📻

[NADRIP]

Hi Chris, the link you provided cannot be accessed without a subscription, so if there is helpful information from that article, can you quote it?  (Or if someone else has a subscription can you post what you believe is useful from this article?)

It seems a bit of semantics to say that Pandora's music analysts DO "qualify the musicological components" but DO NOT "draw those analogies" - how else are the analogies drawn if not by what is input?  Perhaps the algorithm does the actual work of "drawing those analogies" but again what it calculates is entirely dependent on what is input.  There's an old saying, "garbage in / garbage out" and I'm not implying that Pandora's process is "garbage" just that what is input is the ultimate driver of what is calculated by the algorithm.  I think to deny this is to deny a basic fact of how every computer-assisted operation works.  The algorithm is just an "assisting" mechanism, able to make calculations extremely fast and automatically.  But there is no aspect of the process that is not ultimately the responsibility of the human element.  One could note that there are two human inputs, both the input of the "musicological components" and the input of the algorithm code, but from my perspective they are both part of the human element, so if your response is implying a certain thing "x" may not be input by a "music analyst" but instead is input by a coder, I'm grouping all of that together as the human element.

ppaul3 had an excellent observation which related to the algorithm in his response to me in another thread and I believe it's useful to quote it here as well:

"I have two artists names. As Paul Adams I do well in the new age genre w/ around 132 million streams. Last year I decided to push the artistic envelope and release music in a very different genre (Americana). I used a different name - PD Adams - so I wouldn't confuse my followers. When I feature as Paul Adams I usually do OK. But, when I feature as PD Adams I get almost nothing. I am assuming then that 'Feature' is based on the algorhythm recognizing you and seeing that you have alot of streams behind you. In my experience that is the difference. So I guess the name of the game is 'get streamed alot'. It's like the Steve Martin Joke...'how to have a million dolars and never pay taxes ....first ...get a million dollars...'"

It seems clear that something in the human input process causes the algorithm to behave very differently if a song is by someone with an established audience verses by someone without an established audience.

As for the ability to manually change information, of course it can be manually changed, as can any information.  Probably what is meant by that is that once a "music analyst" puts information in, it would take a coder to change it, so it's not policy to make manual changes.

Although all comments are welcome, and potentially useful as each has a different perspective on the process, I'm not sure if anything in the comments above makes anything in my initial post untrue or inaccurate.  So please go back and note if you believe that is the case.  My view is that the elements that seemed like disagreement were because of understanding the human element differently, me looking at it as both the "music analysts" and the coders, and the latest response looking at the human element as only the "music analysts" - that's why i can't find anything in my original post that this latest post seems to suggest is untrue.  Also, posting relevant quotations from the cited article can clarify things so I hope that happens.

[kuttmagicmusic]

I absolutely understand and appreciate your reply.. My listening is and
always has been different than most so like everyone in this world my
listening preferences are based on my current state of mind is all I'm
saying. Sometimes I'd be in a good upbeat mood and the station will keep up
until about the 5th song to play which will change the entire vibe. Maybe
algorithm was the wrong word. If I were on the Pandora team I'd want to
specifically spend my time listening to music and categorizing it into
those types of categories so that when a customer creates their playlist
there are people who would create the same type of playlist and the system
can help someone stay on track with the way they're feeling for that day..
Most would disagree due to someone saying that they don't want to be sad
but just because a song makes your emotions rise doesn't mean you're
depressed sometimes situations like what I'm explaining could help release
much needed tears someone was holding back than boom the next day they're
creating a new positive upbeat inspirational playlist.. If you look at my
profile my numbers are low because I mostly just listen because it's so
hard to go back thumb up a song or add it to a playlist or create a new
playlist I'd create a space where someone could be offered a playlist
similar to listening experiences or suggests other profiles with similar
interests.. I don't know I have a Big brain and think about 1 million
things at once so this might be confusing to someone who usually thinks
inside the box.. Outside the box doesn't exist to most of the world.

[NADRIP]

reply test

[NADRIP]

Excellent example.  Let's break down what the current algorithm is and is not doing.

The current algorithm creates stations based on one station for each artist / band, so that station has to include songs of every kind of mood that artist / band released, and then from the average of the songs on that station, other songs will be added, but again songs of every kind of mood.

What the current algorithm is not doing is what many humans prefer, to create and listen to stations based on mood (ie chill vibes, so sad, in love, happy and upbeat, etc.) or activity (sleep, study beats, road trip, running, backyard bar-b-que, etc.)

So, let's say you're in a happy, upbeat mood and craving the Beatle's version of "Shake It Up, Baby,"  So, you go to the Beatles radio station on Pandora, but instead "Hey Jude" plays, which was perfect for your mood last week but today is a buzzkill.  Then "Strawberry Hills" plays and again it doesn't fit your mood.  Then a bunch of songs from other artists / bands play that also don't fit your mood.

As I said before, this is a weakness of all algorithms, that they are heavily Pareto Principle, and do well in some areas, maybe up to 80%, but don't do well at all in the other areas, at least 20%.  So, what this current algorithm seems to do best in is in musical genres where the mood/activity is usually synonymous with the genre itself, like new age music, or jazz, or even many classical songs.  Someone wanting to listen to new age music probably wants relaxing, background vibes, so all songs from a typical new age artist are going to be in this same general mood / activity group, as are all the songs from other artists / bands that the algorithm pulls.

And what the current algorithm doesn't do well in is the opposite, artists / groups that release songs representing a wide variety or moods or appropriate associated activities.

So, the solution, as I already noted, is a hedging algorithm.  Let's say that another algorithm was coded that would use mostly the same current inputs as the current algorithm, so not much more work to have two algorithms as opposed to just one.  And what this second algorithm would do is create mood-based playlists.  For instance, one possibility is it could create additional radio stations on each artists site, so instead of just the main Beatles radio station, there could be ones like "Beatles - Happy & Upbeat" and "Beatles - So Sad" and "Beatles - in Love" and "Beatles - Chill Vibes".  Alternately, this new algorithm could create just some master mood / activity playlists and put them on a mood / activity page and pull songs from many artists / groups.  Or it could create both.

Now, going back to you craving to hear "Shake It Up, Baby" by the Beatles, you could just go to the "Beatles - Happy & Upbeat" radio station and hear that song and many more that will fit your mood.  Or you could just search for "happy upbeat" and get the generic "Happy & Upbeat" radio station.,

Again, the new algorithm could create these stations automatically, and use mostly the inputs that the first algorithm is already using, so other than the initial coding, this wouldn't be much more ongoing work for Pandora, but would offer a hugely different listening experience, and would with the second algorithm hedge one of the main weak areas that the current algorithm has.

[NADRIP]

I notice I cannot "edit" or 'delete" my comments, like the above, for example.  None of the commands that show with I click on the 3 dots seem to work.  I am using Safari 16.5.

Also, I notice someone chose a reply I made as the "solution" which is inappropriate.  I conclude a moderator must have done this.  Previously, there were several dozen short, inane comments trying to derail the thread like "how are you? i'm fine" and I'm glad these were appropriately cleaned up and deleted.  But, before that, one of these comments was (again, probably by a moderator since I don't know who else has the ability to do this) selected as the "solution".  I don't believe this thread has a solution yet as the purpose of the thread is to have an ongoing discussion of something that is not going to have an easy solution.

[NADRIP · Accepted solution]

I have a couple more observations about the Pandora algorithm.  The first is something that was noted in another thread about feature track.  This comment was given by CatleaMusic, "It doesn't seem like Pandora's marketing programs and algorithms are designed to support or promote new artists."  In support of this, this is how Pandora explained their feature track marketing tool in the latest webinar - if someone has a large number of existing streams, that person will tend to get a proportionally large benefit from feature stream, whereas if someone has almost no existing streams, that person will tend to get almost no benefit from feature stream.  Thus, in this case it seems what the algorithm is mainly doing when feature song is on is adding an X% bonus to the normal chance for the algorithm to play a song, so if it's, say, 10% rounded down, then if someone is already getting 1000 streams a day they might see a benefit of 100 streams a day from feature stream, but if someone is getting 1 stream a day they might not see any benefit from feature stream.  Pandora also reports that getting a lot of thumbs up or thumbs down will affect how often the algorithm plays a song, but the truth is that Pandora plays radio-style music to mostly passive listeners so there isn't a high number of thumbs up or thumbs down for most songs.  Thus, for most songs, it's mostly existing streams driving how this marketing feature works, and the algorithm is actually doing very little beyond adding a simple bonus percentage.

Another thing that has happened is that some of my other songs have been added to Pandora radio, but one thing I've found out is that any time I have a collaborating primary artist or even just another artist who is merely featured on one of my songs, that song goes on the radio of a separate station.  This might be a good thing, because sometimes my collaborations or songs with different featured singers result in song which have unique aspects from my other songs.  I will use the example of my collaboration with Tiago Videira, who mostly composes instrumental tracks, which range from classical music to traditional European regional music.  Whereas my songs which aren't collaborations with him tend towards country, rock, pop (with Oualid Ekami as featured singer), or dream pop (with Nekane as featured singer).  Tiago and I made three songs together, all in an American folk vein that is truly the product of collaboration and distinct from my individual songs or his individual songs.  Interestingly, one of these three songs, "Rollin' On," is a fairly popular song for people who like chill music, though that's on a completely different station yet again because we also used as our featured singer on that song Arkansas native but nowadays Las Vegas-based Justin Carder.  Our other two songs, which are on a combined station together, are more niche.

So, although my songs are all shown on my main page, only my songs made under the NADRIP name with no other collaborating or featured artists are on that page's radio station, and if there is a collaborating and/or featured artist, then there is are separate pages with separate radio stations with those songs.  Or that's my current understanding.  These songs have all gone up just in the past few days so I haven't had a chance to find out more yet.  In AMP, I have to go to each page separately to see the details about songs which are playing on that radio station, so although all my songs are collected on my Pandora artist page, they are not collected on AMP, and AMP treats each collaboration or different featured artist as a different artist name (ie "NADRIP" is different from "NADRIP & Tiago Videira" is different from "NADRIP & Nekane" is different from "NADRIP, Tiago Videira & Justin Carder").

I did, however, see that my song "Coulda Stayed" (which is a song by NADRIP with no collaborating or featured artists) streamed on the station for NADRIP, Tiago Videira & Justin Carder.  So there must be some chance that songs by individual named artists which are similar enough to the collaborative songs will also be included.